⚡ Quick verdict
- C
Best AI model overall → Claude Fable 5.2. New #1 on the Artificial Analysis Intelligence Index at 65 — but slow and pricey per task at max effort.
- G
Best for automation & security → Gemini 4 Carbon. #1 on AutomationBench and CursorBench 4.0, $2/$10 per million tokens until 30 November 2026.
- C
Best model family → GPT-6 (Astra, Sol, Luna). One family from $0.10 to $10 per million input tokens — pick the tier, not the brand.
- C
Best all-rounder → Claude Opus 5.5. Still leads the AA-Briefcase Elo board at 1800, $4/$20 per million tokens.
- K
Best open-weights model → Kimi K3.1 Pro. Full weights on Hugging Face, near-Opus coding scores at half the API price.
- G
Best for the X ecosystem → Grok 4.7. Strong on real-world coding benchmarks, $2/$6 per million tokens — but trails the leaders on the AA Index.
- C
Best cheap everyday model → Claude Haiku 5.5. $0.10/$0.50 per million tokens under 100K-token prompts, AA Index 36.
- M
Best agent volume worker → Muse Ember. Haiku-level scores at $0.14 per task — built for tool calls, routing and extraction at scale.
- D
Best budget batch model → DeepSeek V4.1 Flash. Off-peak pricing halves the bill to $0.15/$0.60 per million tokens.
We may earn a commission if you buy through our links, at no extra cost to you.
The leaderboard changed twice in three days. On 9 November 2026, Anthropic released Claude Fable 5.2 and Meta released Muse Ember on the same day; on 10 November, Google opened Gemini 4 Carbon to everyone. The result is the biggest reshuffle of the AI model rankings this year: Fable 5.2 took first place on the Artificial Analysis Intelligence Index at 65, Carbon took first on three workflow benchmarks, and the cheapest capable models got cheaper still. If you last chose a model in September, your answer is probably out of date.
This is a researched comparison, not a lab test. We ran no prompts through any of these models ourselves. Every score below comes from a named source checked on 10–11 November 2026: the vendors' own launch announcements and pricing pages, Artificial Analysis's Intelligence Index, Vals.ai, CursorBench and the Epoch AI Benchmarking Hub. Vendor-reported numbers are labelled as such throughout, because a benchmark a vendor runs on its own model is a claim, not a verdict. Prices are November 2026 snapshots — pricing changes often in this market, so check the official site before paying. Two warnings before the ranking: Artificial Analysis launched Index v5.0 in late October and regraded part of it in November, so scores from older index versions are not directly comparable with the new ones, and per-token prices mislead — what matters is cost per finished task, because the strongest models burn far more tokens per answer.
If you are choosing between the two biggest chatbot brands rather than the models underneath them, our ChatGPT vs Claude comparison covers the apps, plans and free tiers, and our Claude vs Gemini guide compares the same rivalry from the Google side. Builders who want the cheapest capable model for their own product should also read our best free AI chatbots with no signup roundup, and developers picking a coding model will get more depth from our best AI coding assistants guide.
At a glance: the best AI models in November 2026
| Model | AA Index* | API price per 1M tokens (in/out) | Best for |
|---|---|---|---|
| C | 65 (Max) | $8 / $40 | Hardest coding & research |
| G | ~60 | $2 / $10 (intro, to 30 Nov) | Automation & security |
| C | 57 | $10 / $50 | OpenAI flagship work |
| C | 59 (Max) | $4 / $20 | All-round frontier work |
| K | 55† | $2 / $10 | Open-weights coding |
| G | 35.5 | $2 / $6 | X ecosystem, long coding runs |
| C | 36 (Max) | $0.10 / $0.50 (under 100K tokens) | Cheap everyday work |
| M | 36 (Max) | $0.10 / $0.50 (under 200K input) | Agent volume work |
| D | n/a | $0.30 / $1.20 peak; $0.15 / $0.60 off-peak | Budget batch jobs |
*Artificial Analysis Intelligence Index v5.0.1, read 9–10 November 2026, unless marked. †Kimi's 55 was measured on the older Index v4.3.2 and is not directly comparable with the v5 scores. DeepSeek V4.1 Flash had no published AA Index score in the sources we checked on 11 November. Pricing changes often — treat every figure as a November 2026 snapshot and check the official site before buying.
C
#1 · Claude Fable 5.2 · Best AI model overall
- AA Index
- 65 at Max effort — first of 74 models (v5.0.1)
- API price
- $8 input / $40 output per million tokens
- Best for
- Long, hard technical work: large migrations, long agent runs, research problems
Detail
1. Claude Fable 5.2 — best AI model overall
Anthropic released Claude Fable 5.2 on 9 November 2026, and within a day it sat at the top of the two independent leaderboards that matter most. Artificial Analysis scores it 65 on its Intelligence Index v5.0.1 at maximum effort — first of 74 models, five points clear of Gemini 4 Carbon and six clear of Claude Opus 5.5 — and Vals.ai puts it first too, at 71.3% on its index, ahead of Opus 5.5 at 70.1% and GPT-6 Astra at 65.0%. Anthropic's own launch table has it leading ten of the twelve benchmarks it appears in. The model is aimed squarely at what Anthropic calls "the most daunting technical challenges": large code migrations, long agent runs and research problems, with a 1-million-token context window, up to 128K output tokens, adaptive thinking that is always on, and a training-data cutoff of June 2026. The API model id is claude-fable-5-2.
The price is the surprise in the other direction: Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, 20% less than Fable 5.1's $10/$50 — a flagship that got cheaper as it got better. Anthropic also says it generates output up to 30% faster than Fable 5.1, though no published measurement we found backs that speed claim yet. On coding, the vendor-reported numbers are large: 49.1% on FrontierCode 2.0 at Xhigh effort, up from 32.4% for Fable 5.1 at the same setting, and 45.3% at the lower High setting, close to Opus 5.5's 45.5% at Max. Launched alongside it was Claude Mythos 5.2, the same underlying model with more permissive cyber and biology safeguards, offered only to verified programmes — not something a normal user can buy.
The honest limits are cost and speed, and they are severe. Artificial Analysis measured $26.54 to run one index task with Fable 5.2 at max effort — the highest of the 73 models on that chart — against $9.08 at its High setting (61.82 on the index) and $3.80 at Medium (56.41). It is also slow: 103 tokens per second output, 72nd of 73 models measured, with 44.1 minutes per index task against a 12.9-minute median. And it does not win everything: on Anthropic's own launch table, Gemini 4 Carbon leads AutomationBench at 53.9% against Fable's 51.4%, and Opus 5.5 leads the AA-Briefcase knowledge-work Elo at 1800 against Fable's 1785. For the hardest jobs, Fable 5.2 is the best model in November 2026. For anything routine, running it at max effort is paying first-class prices for an economy ticket.
Claude Fable 5.2
Strengths
- #1 on the Artificial Analysis Intelligence Index (65) and first on Vals.ai (71.3%) — the strongest independent results of any model this month.
- Cheaper per token than its predecessor ($8/$40 vs $10/$50) and cheaper than GPT-6 Astra ($10/$50).
- 1M-token context, 128K max output, adaptive thinking always on — built for long agent runs.
- Big coding jump over Fable 5.1: 49.1% on FrontierCode 2.0 at Xhigh, up from 32.4% (vendor-reported).
Weaknesses
- Most expensive per finished task at max effort: $26.54 per index task, and one of the slowest at 103 tokens per second.
- Loses two rows on Anthropic's own launch table: AutomationBench to Gemini 4 Carbon, AA-Briefcase Elo to Opus 5.5.
- "Up to 30% faster" claim has no published independent measurement yet.
- Mythos 5.2 variant is restricted to vetted organisations — not available to general users.
G
#2 · Gemini 4 Carbon · Best for automation & security
- AA Index
- About 60 (59.4 at High, v5.0) — second overall
- API price
- $2 input / $10 output per million tokens (intro to 30 Nov 2026, then $4 / $20)
- Best for
- Automation workflows, agentic coding, cybersecurity defence work
Detail
2. Gemini 4 Carbon — best for automation & security
Google announced Gemini 4 Carbon on 28 October 2026 but kept it behind a limited rollout to cybersecurity defenders through its Fairwind Program — then, on 10 November, opened it to paying developers, enterprises and consumers via the Gemini API, Google AI Studio and Gemini for business. It is Google's strongest model and the second Gemini 4 model after Argon, built on Argon's foundation with additional post-training for deep reasoning, extended verification and multi-step agent workflows. On the independent numbers it is the closest challenger to Fable 5.2: Artificial Analysis scores it about 60 on the Intelligence Index (59.4 at High), and it ranks #1 on three workflow benchmarks Google cites — CursorBench 4.0 at 60.4%, Perplexity's WANDR, and Zapier's AutomationBench at 53.9%, where it beats Fable 5.2's 51.4%. It is also #2 on the new Multimodal Knowledge Agent benchmark Google built with Artificial Analysis, at more than six times lower average cost per task than the top model there.
The pricing is the best value story among the flagships: $2 per million input tokens and $10 per million output tokens — the same introductory rate as Argon — but only until 30 November 2026, after which it doubles to $4/$20 from 1 December. There is no free tier: Carbon is excluded from the Gemini API free tier, and every processed token is billed, including thinking tokens. The security numbers are the reason it exists: Google reports 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for validating them, 71.0% on CWE-bench for patching at the root cause, and a 0% attack success rate on Gray Swan's indirect prompt-injection benchmark, down from 0.7% for Argon. On internal benchmarks it more than doubles Argon's success rate at generating secure code. The context window is 1,048,576 tokens, with a knowledge cutoff of March 2026.
The honest limits start with the price rise: anyone building on Carbon in November should budget for the December doubling, or the "cheap flagship" story ends on the 1st. Google's own model card also lists hallucinations, possible slowness or timeouts, and greater token consumption at higher thinking levels as limitations, and notes Carbon spends more thinking tokens per turn than Argon at the same per-token price — so the $2/$10 sticker understates the bill on long reasoning jobs. And on the benchmarks Google cited at the first announcement, Carbon sits behind Fable 5.2 overall, which is the usual cherry-picking warning: vendors lead with the tests they win. But for automation, agentic coding and security work at $2/$10, nothing else in November 2026 matches it.
Gemini 4 Carbon
Strengths
- #1 on CursorBench 4.0 (60.4%), WANDR and AutomationBench (53.9%) — the best workflow-benchmark record this month.
- Flagship capability at $2/$10 per million tokens until 30 November — a third of Fable 5.2's price per token.
- Strongest security results published: 88.2% SEC-bench Pro, 71.0% CWE-bench, 0% Gray Swan prompt-injection success (vendor-reported).
- 1M-token context, multimodal input (text, image, video, audio, PDF).
Weaknesses
- Price doubles to $4/$20 from 1 December 2026 — the intro rate is a one-month window.
- No free tier, and every token is billed including thinking tokens; Carbon uses more thinking tokens than Argon.
- Behind Fable 5.2 on the overall Artificial Analysis index (about 60 vs 65).
- Model card lists hallucinations, slowness and timeouts at higher thinking levels.
C
#3 · GPT-6 family (Astra, Sol, Luna) · Best model family
- AA Index
- Astra 57 · GPT-6.1 Sol 54.5 · GPT-6.1 Terra 51 · Luna 27.8
- API price
- Astra $10/$50 · Sol $2/$10 · Terra $1/$5 · Luna $0.10/$0.50 per 1M tokens
- Best for
- Teams already on OpenAI: one family covering flagship work down to bulk jobs
Detail
3. GPT-6 family (Astra, Sol, Luna) — best model family
OpenAI's GPT-6 family arrived in pieces across September and October 2026, and it is best understood as a price ladder rather than a single model. GPT-6 Astra, the flagship, launched 3 September at $10 per million input tokens and $50 output. GPT-6 Sol and GPT-6 Luna followed on 22 September at $2/$10 and $0.10/$0.50. Then GPT-6.1 Sol arrived on 29 September at the same $2/$10 as Sol but with higher scores on every test where both were measured, and GPT-6.1 Terra arrived on 26 October at $1/$5, aimed at visual reasoning, computer use and natural writing. OpenAI's own GPT-6 guide now lists four current models — Astra for the hardest work, GPT-6.1 Sol to balance cost and intelligence, GPT-6.1 Terra for fast visual and writing tasks, and GPT-6 Luna for cost-sensitive volume — and there is no GPT-6 Terra, GPT-6.1 Luna or GPT-6.1 Astra. All eight models in the family share a 1,050,000-token context window and 128,000 maximum output tokens, so price, reasoning settings and tool support separate them, not context. OpenAI calls the Sol and Luna prices permanent, not promotional rates.
On the independent numbers, Astra scores 57 on the Artificial Analysis Intelligence Index, behind Fable 5.2's 65 and Carbon's roughly 60, while GPT-6.1 Sol scores 54.5 — remarkably close to the flagship at a fifth of the price — Terra scores 51.0 and Luna 27.8. OpenAI's own figures put Astra at 74.1% on DeepSWE v1.1 at maximum effort, ahead of GPT-6.1 Sol at 71.9% and GPT-6 Sol at 68.8%. The naming is genuinely confusing: GPT-6 Sol and GPT-6.1 Sol share a $2/$10 list price but are not the same model, and GPT-6.1 Sol is the one OpenAI now recommends. For ChatGPT users, Astra, GPT-6.1 Sol, GPT-6 Sol and Luna are listed in Codex for Plus subscribers at $20 a month, with GPT-6.1 Sol also in ChatGPT Work and Codex but not in Chat, and GPT-6.1 Terra reaching ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users on 26 October.
The honest limits are the ones OpenAI's own pricing creates. Above 272,000 input tokens, the published long-context schedule doubles input and cache rates and raises output rates by 50% for the whole request — Sol becomes $4/$15 and Luna $0.20/$0.75 — so long-context jobs cost more than the headline rates suggest. And the fivefold and twentyfold price steps between Astra, Sol and Luna are price ratios, not quality ratios: Luna at $0.10/$0.50 is not one-fiftieth as capable as Astra, but it is also not a substitute for it — OpenAI positions Luna for routine extraction and summarisation, Sol for recurring coding and agent work, and Astra for the hardest jobs. For a team already on OpenAI, the family is the pragmatic pick: route each job to the cheapest tier that can do it, and the bill drops without changing providers. For raw capability alone, Fable 5.2 and Carbon are ahead.
GPT-6 family (Astra, Sol, Luna)
Strengths
- A complete price ladder in one family: $0.10 to $10 per million input tokens, so every job gets the cheapest tier that can do it.
- GPT-6.1 Sol scores 54.5 on the AA Index at a fifth of Astra's price — the best value flagship-class model from OpenAI.
- Permanent prices, not launch promotions, according to OpenAI.
- Deepest ecosystem: ChatGPT, Codex, and the widest third-party tool support of any model family here.
Weaknesses
- Astra trails Fable 5.2 and Carbon on the independent index (57 vs 65 and about 60) while costing the most per token.
- Confusing names: GPT-6 Sol and GPT-6.1 Sol share a price but are different models; GPT-6.1 Sol is the current pick.
- Long-context surcharge above 272K input tokens doubles input rates for the whole request.
- Luna's 27.8 AA Index score is far behind the other cheap models — Haiku 5.5 and Ember both score 36 at the same price.
C
#4 · Claude Opus 5.5 · Best all-rounder
- AA Index
- 59 at Max effort (v5.0.1) — third overall
- API price
- $4 input / $20 output per million tokens
- Best for
- Serious work where Fable 5.2's cost and slowness are not justified
Detail
4. Claude Opus 5.5 — best all-rounder
Claude Opus 5.5 launched on 22 September 2026 — the same day OpenAI shipped GPT-6 Sol and Luna — and for seven weeks it was the model to beat. Fable 5.2 has now passed it on the overall index, but Opus 5.5 is far from finished: it scores 59 on the Artificial Analysis Intelligence Index v5.0.1 at Max effort, third overall, at $10.04 per index task against Fable 5.2's $26.54 at Max. It still leads the AA-Briefcase knowledge-work Elo board at 1800, ahead of Fable 5.2's 1785 on Anthropic's own launch table, and it scores 70.1% on the Vals.ai index, fourth overall. On Terminal-Bench 4.0 it scores 66.4% in Moonshot's vendor-reported table — a fraction ahead of Kimi K3.1 Pro's 65.7% and well ahead of GPT-6 Astra's 58.2%. Anthropic still advises starting with Opus 5.5, even after launching Fable 5.2 — a telling detail about which model it considers the safer default.
The price is the argument for Opus 5.5: $4 per million input tokens and $20 output — half of Fable 5.2's rates, a fifth of Astra's output price — with $0.20 cached input. For most professional work, the gap between 59 and 65 on the index is invisible in the output and very visible in the invoice. Opus 5.5 is also the reference model every other vendor benchmarks against: Moonshot, Google and OpenAI all put it in their launch tables, which means its scores are the most cross-checked of any model here. It is available through the Claude API, Claude apps and the usual cloud providers, and Claude Pro costs $20 a month for the consumer app.
The honest limits are that it is now second-best inside its own family. Fable 5.2 beats it on most of Anthropic's launch benchmarks, on the overall index and on Vals.ai, and it costs more per token than the new mid-tier models — GPT-6.1 Sol and Kimi K3.1 Pro both charge $2/$10 and score within a few points of it. The practical question is not whether Opus 5.5 is good — it clearly is — but whether its $4/$20 price still earns its place now that $2/$10 models score nearly as well and Fable 5.2 exists for the jobs that truly need the ceiling. For teams that standardised on Opus 5.5 in September and October, the answer is to stay: it remains a top-three model, and switching costs more than the benchmark gap is worth. For new projects, test GPT-6.1 Sol and Kimi K3.1 Pro first.
Claude Opus 5.5
Strengths
- Still #1 on the AA-Briefcase knowledge-work Elo at 1800, ahead of Fable 5.2 — the best model for professional document work on that board.
- Third overall on the AA Index (59) at less than half Fable 5.2's cost per task ($10.04 vs $26.54 at Max).
- Anthropic still recommends starting with Opus 5.5 — the safest default in the Claude family.
- The most cross-checked model here: every rival benchmarks against it.
Weaknesses
- Now second-best inside its own family — Fable 5.2 leads it on most benchmarks.
- $4/$20 per million tokens is double the $2/$10 mid-tier models that score nearly as well.
- No longer the index leader, so new projects should test the cheaper alternatives first.
K
#5 · Kimi K3.1 Pro · Best open-weights model
- AA Index
- 55 (v4.3.2) — fourth of 215 models on the older index
- API price
- $2 input / $10 output per million tokens
- Best for
- Self-hosting, open-weights coding agents, teams that want to own the model
Detail
5. Kimi K3.1 Pro — best open-weights model
Moonshot AI announced Kimi K3.1 Pro on 23 October 2026 and made a promise with it: the full model weights would ship by 4 November. They did — on Hugging Face on 4 November, on schedule, in the repository moonshotai/Kimi-K3.1-Pro, under the Kimi K3.1-Pro License. That matters because a launch is a claim and open weights are a way to check it: anyone can now run, inspect and fine-tune the model. The scale is enormous for an open-weights release — a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters per token, a 1-million-token context window, and a MoonViT-V2 vision encoder. It is available now on Kimi.ai, Kimi Work, Kimi Code and the Kimi API, and the API costs $2 per million input tokens and $10 output, with cached input at $0.20 — exactly half of Claude Opus 5.5's price, and a third less than its predecessor Kimi K3's $3/$15.
On the numbers, Kimi K3.1 Pro is the strongest open-weights model ever measured on the Artificial Analysis index: 55 on the older Index v4.3.2, fourth of 215 models, three points behind Opus 5.5 and nine points above the previous best open-weights model. On Terminal-Bench 4.0, Moonshot reports 65.7% against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra, run at maximum reasoning effort — a gap of 0.7 points on real terminal tasks. The hallucination numbers are the biggest claimed improvement over Kimi K3: Moonshot's internal Kimi Hallu Bench Hard puts the claim-level hallucination rate at 8.3%, down from 23.4%, though that is an internal benchmark and should be treated as direction rather than proof. Independent coding checks are also strong: Requesty's coding benchmark gives it an 81% solve rate against 92% for Opus 5.5 and 67% for GLM-5.3, at $1.04 per solve.
The honest limits start with the licence: it is not a standard open-source licence, and the Kimi K3 licence required a separate agreement for model-as-a-service businesses above $20 million in revenue over 12 months — check the K3.1-Pro License terms before building a business on it. Then there is the hardware: the download is about 1.56 TB, and you need well over 1.5 TB of fast memory just to hold the weights, so self-hosting is a data-centre job, not a home-server project. Moonshot's own benchmarks are vendor-run, mostly inside the Claude Code harness, and one independent audit (R40 Research, 3 November) flagged benchmark-maximisation concerns in its Terminal-Bench and HLE runs — the strongest documented case of that practice this year. And the consumer side has a real complaint: users on Kimi's $99–$100-a-month plans report hitting the 5-hour usage window in under an hour. For teams that want to own their model, though, nothing else in November 2026 comes close.
Kimi K3.1 Pro
Strengths
- Full weights published on Hugging Face on 4 November, as promised — the strongest open-weights model ever measured (AA Index 55).
- Near-Opus coding scores (Terminal-Bench 4.0 65.7% vs 66.4%) at half the API price ($2/$10 vs $4/$20).
- 2.8T MoE, 104B active, 1M context — frontier scale you can actually download and run.
- Hallucination rate cut from 23.4% to 8.3% on Moonshot's internal hard benchmark (vendor-reported).
Weaknesses
- Not a standard open-source licence — check the K3.1-Pro License before commercial serving.
- Self-hosting needs about 1.56 TB of weights and well over 1.5 TB of fast memory.
- Vendor benchmarks flagged for possible benchmark-maximisation by an independent audit (R40 Research, 3 November).
- Consumer subscription complaints: 5-hour usage windows reportedly exhausted in under an hour on $99–$100 plans.
G
#6 · Grok 4.7 · Best for the X ecosystem
- AA Index
- 35.5 — 16th of 54 models on that board
- API price
- $2 input / $6 output per million tokens ($0.50 cached input)
- Best for
- Long-running software engineering tasks, X and Grok users
Detail
6. Grok 4.7 — best for the X ecosystem
xAI released Grok 4.7 on 21 September 2026, built on a larger base model than Grok 4.6 with a longer reinforcement-learning run, and describes it as its most capable model for coding and knowledge work. It is the odd one out in this ranking: its overall Artificial Analysis Intelligence Index score is 35.5, 16th of 54 models on that board — far behind the leaders — yet on specific coding benchmarks it looks like a different model entirely. xAI reports 71.0% on DeepSWE v1.1 at high effort, 64.0% on EEBench, and 37.6% on Terminal-Bench 4.0 at xHigh effort, nearly double Grok 4.6's 20.3%. An independent Epoch AI run puts it at 92.7% on GPQA Diamond and 98.1% on the OTIS Mock AIME, and on LMArena it scores 1442 on text and 1640 on WebDev. The API price is $2 per million input tokens and $6 output, with cached input at $0.50, and the context window is 500,000 tokens — half the 1M window the other flagships now share.
The honest reading is that Grok 4.7 is a specialist, not a generalist. The AA Index combines nine evaluations, and Grok's weak spots drag its average down even where its coding is strong; the Terminal-Bench jump from 20.3% to 37.6% is real progress on agentic terminal work, but 37.6% still trails Opus 5.5's 66.4% and Kimi K3.1 Pro's 65.7% by a wide margin. On writing, ToneBench ranks it 28th of 87 models through its best setting (medium), with a writing score of 79.6 out of 100 — respectable, not leading. What Grok 4.7 genuinely offers is the xAI ecosystem: it is the model inside X and the Grok app, where X Premium costs $8 a month, and it is the model xAI's own products are built on. For a developer choosing purely on benchmarks, the $2/$10 models above it — GPT-6.1 Sol, Kimi K3.1 Pro, Gemini 4 Carbon — score higher for the same money or less. For an X user who wants the best model inside the app they already use, Grok 4.7 is the only answer, and it is a large step up from Grok 4.6.
Grok 4.7
Strengths
- Strong on real-world coding benchmarks: 71.0% DeepSWE (xAI-reported), 92.7% GPQA Diamond (Epoch AI independent run).
- Nearly doubled Grok 4.6 on Terminal-Bench 4.0 (37.6% vs 20.3%, vendor-reported).
- Cheapest flagship output price here after the budget tier: $6 per million output tokens.
- Native to X and the Grok app — no new account or workflow for X users.
Weaknesses
- Weakest overall AA Index score of the flagships here (35.5) — a generalist gap, not just a rounding error.
- Terminal-Bench 4.0 still trails Opus 5.5 (66.4%) and Kimi K3.1 Pro (65.7%) by nearly 30 points.
- 500K context window is half the 1M the other flagships offer.
- xAI published no new model in November as of 1 November — Grok 4.7 is a September model in a November ranking.
C
#7 · Claude Haiku 5.5 · Best cheap everyday model
- AA Index
- 36 at Max effort (v5.0.1) — best cheap-model score
- API price
- $0.10 input / $0.50 output per million tokens (up to 100K-token prompts)
- Best for
- High-volume everyday work: summarising, drafting, classification, chat
Detail
7. Claude Haiku 5.5 — best cheap everyday model
Anthropic released Claude Haiku 5.5 on 7 October 2026, and it quietly changed the economics of everyday AI work. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output — 90% below Haiku 4.5's $1/$5, and the same headline rate as OpenAI's GPT-6 Luna. Anthropic says it is about 75% cheaper to run than Haiku 4.5 on average, after allowing for its higher token use. On capability it leads its price class clearly: 36 on the Artificial Analysis Intelligence Index v5.0.1 at max effort, against 28 for GPT-6 Luna and 31 for Gemini 3.8 Flash, and 58.73% on the Vals AI index against 52.32% for Luna. Anthropic's own launch table has it at 1620 on GDPval-AA v2.1 against Luna's 1437, 72.4% on OSWorld 2.1 against 48.9%, and 39.2% on Terminal-Bench 4.0 against 16.4%. It has a 1-million-token context window and up to 128K output tokens, and it is available through the Claude API, AWS, Google Cloud, Microsoft Foundry and GitHub Copilot.
The price has a trap, and it is the most important detail in this section: above 100,000 tokens, Anthropic bills the whole request at five times the base rates — $0.50 input and $2.50 output. A prompt that crosses the line by one token costs five times as much as the same prompt just under it. That makes Haiku 5.5 the best cheap model for short prompts and a different, much pricier model for long ones — and Artificial Analysis's published cost figures did not initially model the 5x step, so some cost-per-task comparisons understate it. The model is also more willing to say it does not know than its rivals: on AA-Omniscience it answers 36% of factual questions correctly against 55% for Gemini 3.8 Flash, with a 40% hallucination rate, partly because it declines questions rather than guessing. And one early AutomationBench score was depressed by a pre-release safety bug that made the model over-refuse, which Anthropic was working to fix.
For high-volume work — customer support drafts, document summaries, classification, chat — Haiku 5.5 is the default pick in November 2026: the highest-scoring model at the lowest price, with the Claude family's writing quality. The rule is simple: keep prompts under 100,000 tokens, or route the long ones to GPT-6 Luna, which keeps its base rates up to 272,000 input tokens. Anyone building an agent that processes long documents should set that routing rule on day one, before the first invoice arrives.
Claude Haiku 5.5
Strengths
- Best cheap-model score on the AA Index (36) — ahead of GPT-6 Luna (28) and Gemini 3.8 Flash (31) at the same $0.10/$0.50 headline price.
- 90% cheaper than Haiku 4.5; about 75% cheaper to run on average (Anthropic-reported).
- Strong on computer-use tasks: 72.4% on OSWorld 2.1 against Luna's 48.9% (Anthropic-reported).
- 1M-token context, widely available across clouds and GitHub Copilot.
Weaknesses
- Price jumps 5x above 100,000-token prompts ($0.50/$2.50) — a cliff, not a slope.
- Weak factual recall: 36% on AA-Omniscience against 55% for Gemini 3.8 Flash; declines questions rather than answering.
- Early AutomationBench score was depressed by a safety bug that caused over-refusal.
- Some published cost-per-task figures do not model the 5x long-prompt step.
M
#8 · Muse Ember · Best agent volume worker
- AA Index
- 36 at Max effort — level with Claude Haiku 5.5
- API price
- $0.10 input / $0.50 output per million tokens (up to 200K input)
- Best for
- Tool calls, routing, extraction and coding subtasks inside agents, at scale
Detail
8. Muse Ember — best agent volume worker
Meta released Muse Ember on 9 November 2026 — the same day as Claude Fable 5.2 — and it is the opposite kind of launch: not a flagship, but the smallest and cheapest model in Meta's Muse family, built for the work agents repeat all day. It costs $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens, for requests up to 200,000 input tokens (above that, $0.30/$1.50). It lives on the Meta Model API at api.meta.ai and in Muse Code as model id muse-ember-1.0, with a 1-million-token context window and up to 512,000 output tokens, taking text, images, audio and video as input. Meta says it is the first model trained using Muse Wonder Full, its largest model, and pitches the pair as a team: "Muse Wonder plans, Muse Ember handles the volume." Early partners named on the launch page include Box, Cline, Oracle, Cloudflare and You.com, and it is also listed on OpenRouter and several gateways.
The headline claim is value, and the independent numbers back it: Ember scores 36 on the Artificial Analysis Intelligence Index at max effort — level with Claude Haiku 5.5's best score — at $0.14 per index task against Haiku's $0.41, about a third of the cost. Meta's launch post puts Muse Spark 1.3 at 37 on the same index at $2.39 per task, so Ember delivers nearly the same score as its bigger sibling for a seventeenth of the cost per task. It is a reasoning model with no off switch, and it is designed for sub-agent tasks, tool calls, routing, extraction, coding subtasks and long document parsing — the small jobs an agent repeats hundreds of times a day, where per-call cost matters more than peak capability.
The honest limits are that Ember is a worker, not a thinker. Artificial Analysis found knowledge is its weak spot: it declines 64% of trivia-style questions, so it is the wrong model for factual question-answering without retrieval. It cannot be downloaded — it runs only through Meta's API and services that resell it, unlike Meta's open-weight Muse Glimmer and Muse Spark Open. And the $0.10/$0.50 rate only holds up to 200,000 input tokens; above that the price triples. For agent builders, the Ember pattern is the real story of November 2026: use the expensive model to plan, and the cheap model to execute. Pair Ember with Muse Wonder for planning, or with Fable 5.2 or Opus 5.5 for the hard steps, and the volume work costs almost nothing.
Muse Ember
Strengths
- Haiku-level scores (AA Index 36) at $0.14 per task — about a third of Haiku 5.5's $0.41.
- Cheapest cached input here: $0.004 per million tokens.
- 1M-token context, up to 512K output, multimodal input (text, image, audio, video).
- First model trained using Muse Wonder Full — designed as the volume worker under a planning model.
Weaknesses
- Weak factual knowledge: declines 64% of trivia-style questions (Artificial Analysis).
- Hosted only — no downloadable weights, unlike Meta's open-weight Muse models.
- Price triples above 200,000 input tokens ($0.30/$1.50).
- New model (9 November) with no track record beyond launch-week benchmarks.
D
#9 · DeepSeek V4.1 Flash · Best budget batch model
- AA Index
- No published score in the sources checked on 11 November
- API price
- $0.30 input / $1.20 output peak; $0.15 / $0.60 off-peak per 1M tokens
- Best for
- Batch jobs, classification and coding tasks you can schedule off-peak
Detail
9. DeepSeek V4.1 Flash — best budget batch model
DeepSeek V4.1 Flash first went live on 10 September 2026, with the current V4.1-Flash-1017 revision arriving on 17 October. It is a 552-billion-parameter mixture-of-experts model with a 1-million-token context window and up to 384,000 output tokens, and its pricing is the most unusual in this ranking: DeepSeek charges by the hour of the day. At peak hours it costs $0.30 per million cache-miss input tokens and $1.20 output; off-peak, both halve to $0.15 and $0.60, with cache-hit input at $0.006 peak and $0.003 off-peak. Peak hours are weekday mornings and afternoons in Beijing time, excluding Chinese public holidays. The DeepSeek chat app itself is free, and the API is the cheapest way to run a capable model in bulk if you can schedule the work.
On benchmarks, the Epoch AI figures are strong for the price: 43.9% on FrontierCode at high effort (against 18.8% for the older V4 Flash), 89.8% on GPQA Diamond, 72.9% on ARC-AGI-2, 67.4% on FrontierMath tiers 1–3 and 98.3% on the OTIS Mock AIME. Tencent's own comparison, which reran 34 benchmarks, found DeepSeek V4.1 Flash winning 7 of them against its Hy4 model — concentrated in coding and automation, including DeepSWE, AutomationBench and its own Foundation Model Bench — though those scores were compiled by Tencent, not an independent lab. One independent measurement is telling: on ToneBench's writing benchmark, DeepSeek V4.1 Flash at max effort ranks higher than Grok 4.7 at medium for $0.017 per script against $0.215.
The honest limits are the ones that come with the price. The off-peak discount only helps if your workload can wait: a real-time product cannot schedule its users' requests for off-peak hours. DeepSeek publishes no USD list price for this model in some places, so third-party trackers quote platform rates — the $0.15/$0.60 figures are gateway rates read from the live rate config on 11 November, not a DeepSeek-published USD list. And we found no published Artificial Analysis Intelligence Index score for it in the sources checked on 11 November, which makes it harder to compare head-to-head with the models above. For batch work — tagging, classifying, cleaning, bulk coding tasks — scheduled off-peak, it is the cheapest capable model in this ranking. For anything live, Haiku 5.5 or Ember at the same $0.10/$0.50 flat rate are simpler.
DeepSeek V4.1 Flash
Strengths
- Cheapest capable model for batch work: $0.15/$0.60 per million tokens off-peak.
- Strong benchmark record for the price: 43.9% FrontierCode, 72.9% ARC-AGI-2, 98.3% AIME (Epoch AI).
- 1M-token context, 384K max output, native multimodal input.
- Free chat app; API available through DeepSeek and third-party gateways.
Weaknesses
- Off-peak discount only helps schedulable batch work — useless for real-time products.
- No published Artificial Analysis Intelligence Index score in the sources we checked, so head-to-head comparison is harder.
- No DeepSeek-published USD list price in some channels — figures are gateway/platform rates.
- Peak-hour pricing doubles the bill if your jobs land in Beijing business hours.
Pricing changes often — treat every figure above as a snapshot and check the official site before buying. Free-tier credit counts and commercial-use terms change frequently.
How to choose the right AI model in November 2026
The models above are only useful inside a workflow, and the cheapest way to choose is to match the model to the job before you look at the price. Here is the method that works, in the order that matters.
1. Sort your work into three buckets. Hard jobs — long coding tasks, big migrations, research that takes hours — go to Fable 5.2 or Gemini 4 Carbon. Everyday jobs — drafting, summarising, support replies, classification — go to Haiku 5.5, Muse Ember or GPT-6 Luna. Volume jobs inside an agent — tool calls, routing, extraction — go to Ember or Haiku 5.5. Most teams discover that 90% of their calls belong in buckets two and three, and that the flagship only earns its price on the remaining 10%. Our best AI SDR tools guide shows the same bucketing method applied to sales outreach, and our best AI SEO tools roundup shows which models the SEO tools in it are actually built on.
2. Compare cost per finished task, not cost per token. Fable 5.2 at max effort costs $26.54 per Artificial Analysis index task; at High it costs $9.08 for a 61.82 score instead of 65. Muse Ember costs $0.14 per task for a 36. The per-token price tells you almost nothing about the bill. Run your own representative job through two or three candidate models, count the tokens each one actually uses, and multiply by the rate card. A model that is cheap per token and verbose per answer can cost more than an expensive model that answers in fewer tokens.
3. Watch the price cliffs. Three models in this ranking have one: Haiku 5.5 jumps 5x above 100,000-token prompts, Muse Ember triples above 200,000 input tokens, and GPT-6 models double input rates above 272,000 tokens. Gemini 4 Carbon doubles on 1 December 2026. Put the cliff in your routing logic before you ship, not after the first invoice. If your prompts regularly cross a cliff, pick the model whose cliff is furthest away.
4. Treat vendor benchmarks as claims, not verdicts. Every vendor leads with the tests it wins — Google with AutomationBench, Anthropic with FrontierCode, Moonshot with Terminal-Bench. The independent numbers (Artificial Analysis, Vals.ai, CursorBench, Epoch AI) are the ones to weight heaviest, and even those change when the index version changes. The only benchmark that settles the question is your own: a fixed set of your real tasks, scored by your team, run on each candidate model.
How we ranked them
Since this is a researched comparison rather than a lab test, here is exactly what goes into the ranking: (1) independent capability — the Artificial Analysis Intelligence Index v5.0.1, Vals.ai and CursorBench scores, checked 9–11 November 2026; (2) verification — is the model live, available and checkable on its vendor's own pages in November 2026; (3) value — cost per finished task, not per token, including price cliffs and intro-rate expiries; (4) honesty — does the vendor publish its limits and prices clearly, or hide them behind a sales page. Vendor-reported benchmark scores are labelled wherever they appear and count for less than independent ones. All rankings are a snapshot of November 2026 — this market moves in days, not quarters, so treat them as guidance, not gospel.
Which AI model should you pick?
- Hardest coding or research, money no object → Claude Fable 5.2. AA Index 65, first overall — $8/$40 per million tokens, and slow at max effort.
- Automation, agentic coding or security work → Gemini 4 Carbon. #1 on AutomationBench and CursorBench 4.0 — $2/$10 until 30 November, then $4/$20.
- Already on OpenAI → GPT-6 family. Astra for the hardest jobs, GPT-6.1 Sol ($2/$10) for serious coding, Luna ($0.10/$0.50) for bulk work.
- Best all-rounder → Claude Opus 5.5. AA Index 59, still #1 on the AA-Briefcase Elo — $4/$20 per million tokens.
- Want to own the model → Kimi K3.1 Pro. Open weights on Hugging Face, near-Opus coding at half the API price — $2/$10 per million tokens.
- X or Grok user → Grok 4.7. Strong coding benchmarks inside the X ecosystem — $2/$6 per million tokens.
- Everyday work at scale → Claude Haiku 5.5. AA Index 36 at $0.10/$0.50 — keep prompts under 100,000 tokens to avoid the 5x jump.
- Agent volume work → Muse Ember. Haiku-level scores at $0.14 per task — the cheapest way to run tool calls at scale.
- Batch jobs you can schedule → DeepSeek V4.1 Flash. $0.15/$0.60 off-peak — the cheapest capable model for work that can wait.
Still deciding? Start with the free tiers — ChatGPT, Claude, Gemini and DeepSeek all have free plans for trying the models in the apps — then check current AI deals before paying for anything, and browse our best AI resume builders guide for a worked example of how the same model-choice logic applies to a single use case. If you are comparing the chatbot apps rather than the models, our ChatGPT vs Claude free tier comparison covers what you actually get for $0.
Frequently asked questions
What is the best AI model in November 2026?
On the evidence available on 11 November 2026, Claude Fable 5.2 is the best AI model overall: it scores 65 on Artificial Analysis's Intelligence Index v5.0.1, first of 74 models, ahead of Gemini 4 Carbon at about 60 and Claude Opus 5.5 at 59. It also leads Vals.ai's index at 71.3%. The caveat is cost and speed: at maximum effort it is the most expensive model per task Artificial Analysis measured ($26.54) and one of the slowest (103 tokens per second). If your work is long, hard coding or research, it is the pick; for everyday work, the cheaper models below are the smarter buy.
Is Claude Fable 5.2 better than GPT-6 Astra?
On the independent benchmarks we checked, yes — Fable 5.2 scores 65 on Artificial Analysis's Intelligence Index v5.0.1 against 57 for GPT-6 Astra, and 71.3% against 65.0% on the Vals.ai index. Fable 5.2 is also cheaper per token than Astra ($8/$40 per million input/output tokens against Astra's $10/$50). What Astra still wins is the ecosystem: it is the flagship inside ChatGPT and OpenAI's API, so if your team, tools and workflows already run on OpenAI, switching costs may matter more than the benchmark gap. Check the official pricing pages before buying — pricing changes often.
What is the cheapest good AI model in November 2026?
Three models share the same headline rate of $0.10 per million input tokens and $0.50 per million output tokens: Claude Haiku 5.5 (for prompts up to 100,000 tokens — above that it jumps fivefold to $0.50/$2.50), GPT-6 Luna, and Meta's Muse Ember (up to 200,000 input tokens, then $0.30/$1.50). On Artificial Analysis's Intelligence Index, Haiku 5.5 and Muse Ember both score 36, well ahead of Luna at 28, so Haiku 5.5 and Ember are the better cheap picks on capability. DeepSeek V4.1 Flash is also cheap for batch work: $0.30/$1.20 at peak hours and half that off-peak.
Is Kimi K3.1 Pro really open source?
Not open source in the usual sense — open weights is the accurate description. Moonshot AI published the full Kimi K3.1 Pro weights on Hugging Face on 4 November 2026, as promised at launch, under its own Kimi K3.1-Pro License, a Modified-MIT-style text that permits commercial use and serving. You can download and run the model yourself, but the download is about 1.56 TB and you need well over 1.5 TB of fast memory just to hold the weights, so for most people the $2/$10 per million token API is the practical route. Check the model card before planning a self-hosted deployment.
Which AI model is best for coding in November 2026?
For the hardest coding work, Claude Fable 5.2: Anthropic reports 49.1% on FrontierCode 2.0 at Xhigh effort, up from 32.4% for Fable 5.1, and it leads the Artificial Analysis index. For day-to-day coding inside an editor, Gemini 4 Carbon ranks first on CursorBench 4.0 at 60.4%, and Claude Opus 5.5 scores 66.4% on Terminal-Bench 4.0 in Moonshot's vendor-reported table. For coding on a budget, GPT-6.1 Sol at $2/$10 per million tokens or Kimi K3.1 Pro at the same price both score close to the flagships on agentic coding benchmarks. All of these figures except the Artificial Analysis and CursorBench scores are vendor-reported — treat them as direction, not proof, and run your own test on your own codebase.
Do I need the most expensive AI model?
Almost certainly not for most work. The per-token price is the wrong number to compare — what matters is cost per finished task, and the most capable models use far more tokens per task. Claude Fable 5.2 at maximum effort costs $26.54 per Artificial Analysis index task, against $9.08 at its High setting and $0.14 for Muse Ember. The practical rule is to match the model to the job: Haiku 5.5, Muse Ember or GPT-6 Luna for high-volume routine work, GPT-6.1 Sol or Kimi K3.1 Pro for serious coding at a fair price, and Fable 5.2 or Gemini 4 Carbon only for the hardest jobs where a better answer is worth waiting and paying for.