AI Model Benchmarks, Prices and Speed, Tracked

Theme hub · AI models · Updated 21 NOVEMBER 2026

How to read this page. Scores marked lab-reported are the lab’s own results. Third-party scores name who measured them. Every figure carries its date; prices are list prices per million tokens as shown on each lab’s own page on the date given, and they change often.

$8 / $40Top list price per 1M tokens (Fable 5.2 and GPT-6.2 Astra, 21 Nov)
174.17GPT-6.2 Astra on Epoch’s index; Fable 5.2 173.5 (17 Nov)
1,448Fable 5.2 (Max), Arena Agent Arena Index (18 Nov)
100%GPT-6.2 Astra on ARC-AGI-3, Provider Adapter harness (ARC Prize, 17 Nov)
210 tok/sInkling 2 Mini on a Mac Studio (according to EXO Labs, 21 Nov)
3.61 moDoubling time of agent task length (METR, 20 Nov)

The short answer: at the top, Anthropic and OpenAI are level. Claude Fable 5.2 and GPT-6.2 Astra both list at $8 per million input tokens and $40 per million output tokens, sit 0.67 points apart on Epoch AI’s capability index, and swap the lead depending on the test: Fable 5.2 leads Arena’s agent index and the Artificial Analysis index, GPT-6.2 Astra reached 100% on ARC-AGI-3 with the harness that keeps its reasoning state. Google’s Gemini 4 line is cheaper per task: Arena puts Gemini 4 Neon second on its agent index at a median $2.82 a task. Below the frontier, list prices fall to $2.1 per million output tokens for Mistral Large 4, and open models now run on a Mac: according to EXO Labs, Inkling 2 Mini writes 210 tokens a second on a Mac Studio. This page tracks the scores, the prices and the speed, each one dated and sourced.

Latest

  • 21 Nov
    A laptop-sized model, at desktop speed. According to EXO Labs, Thinking Machines’ open-weights Inkling 2 Mini runs at 210 tokens a second on a 96GB M5 Ultra Mac Studio and 125 on a 48GB M5 Max MacBook Pro; EXO says it built a weather app 1.8x faster than Claude Sonnet 5.5 and 11x cheaper at API pricing, in a test whose method it has not published. Our story
  • 20 Nov
    Gemini 4 Neon is second on Arena’s agent index. Arena ranked Neon #2 with 1,336 points at a median $2.82 per task, and said it scores higher than #3 Claude Opus 5.5 (Max) for 53% less per task. Our story
  • 20 Nov
    METR: agent task length doubles every 3.61 months. Its harder test suite puts frontier models at up to 27 hours on verifiable tasks, with weaker judgment than human experts where feedback is unclear. Our story
  • 17 Nov
    GPT-6.2 Astra: $8/$40 and 100% on ARC-AGI-3. OpenAI priced it 20% below GPT-6 Astra, level with Claude Fable 5.2; ARC Prize verified it as the first model to 100% on ARC-AGI-3 with its Provider Adapter harness. Our story
API list price per million output tokens on 21 Nov 2026: Claude Fable 5.2 $40, GPT-6.2 Astra $40, Gemini 4 Neon $15, Gemini 4 Carbon $10, Grok 4.7 $6, Qwen4-Max $4.50, DeepSeek V4 Pro $3.96 at peak, Mistral Large 4 $2.10
Output is the larger cost for most agent work. Gemini 4 Carbon moves to $20 on 1 December. Per-token prices are not cost per task: see what benchmarks tell a buyer, below.

Frontier models: the latest flagship from each lab

The newest top model from each major lab and from the other labs our newsroom covered in the last 60 days, with the headline scores the lab itself published. Where a lab published no table, we say so.

  • Anthropic
    Claude Fable 5.2
    Launched 9 Nov 2026

    49.1% on FrontierCode 2.0 agentic coding at Xhigh (Opus 5.5: 45.5%); 41.3% on the Remote Labor Index (Fable 5.1: 17.9%) lab-reported. Anthropic still recommends starting most coding and knowledge work on Claude Opus 5.5. Our story

    $8 / $40 per 1M tokens
  • OpenAI
    GPT-6.2 Astra
    Released 17 Nov 2026

    On Agents’ Last Exam, 5.5 points ahead of Claude Fable 5.2 “at just 5% of the cost”; GDPval v2 up 6.2 points on GPT-6 Astra with 22% fewer output tokens; 84.5% on OpenAI’s new MedChemBench lab-reported. Our story

    $8 / $40 per 1M tokens
  • Google
    Gemini 4 Carbon and Gemini 4 Neon
    Carbon 10 Nov, Neon 18 Nov 2026

    Carbon: #1 on CursorBench 4.0, WANDR and AutomationBench at more than 2X lower cost per task than the next best external model. Neon: 61 on the Artificial Analysis Intelligence Index (#3), 70.9% on the Vals Index (#2), 58.6 on the AA Coding Agent Index (#1) lab-reported. Our story

    Carbon $2 / $10 ($4 / $20 from 1 Dec); Neon $3 / $15 per 1M tokens
  • xAI
    Grok 4.7
    xAI’s current flagship

    xAI’s models page calls it “Our flagship model for code and everything else”, with a 500k-token context. The page lists no benchmark scores, and our newsroom has not reported a launch table. xAI models page

    $2 / $6 per 1M tokens
  • Alibaba (Qwen)
    Qwen4-Max
    Launched 16 Nov 2026 (blog 15 Nov)

    75.7 on DeepSWE v1.1 (Opus 5.5: 74.2) but 48.5 on Terminal-Bench 4.0 (Opus 5.5: 59.6); top score in 4 of 16 panels in Alibaba’s own chart lab-reported. Not open weight. Our story

    $1.5 / $4.5 per 1M tokens
  • DeepSeek
    DeepSeek V4 Pro and V4.1 Flash
    Model versions V4-Pro-0813 and V4.1-Flash-1017

    We have not traced a DeepSeek-published benchmark table for these versions. Third party: Vals AI scored V4.1 Flash 1017 at 16.3 of 21 objectives on its offensive-cyber task, against 15.7 for GPT-6 Sol Daybreak. Our story

    Pro $1.32 / $3.96 peak, half off-peak per 1M tokens
  • Mistral
    Mistral Large 4
    Weights released 29 Oct 2026

    Open-weight, 1.05T total parameters and 52B active. Mistral says it ranks #1 on Cognition’s FrontierCode among open models lab-reported. Model-as-a-Service hosts need a separate licence. Our story

    $0.7 / $2.1 per 1M tokens
  • Meta
    Muse Spark Open; Muse Ember
    Spark weights 15 Oct, Ember 9 Nov 2026

    Muse Spark Open: 816B parameters, 42B active, 1M-token context. Muse Ember is Meta’s smallest model, built for agent sub-tasks. We have not traced a Meta benchmark table. Our story

    Ember $0.10 / $0.50 per 1M tokens
  • Moonshot
    Kimi K3.1 Pro
    Weights released 4 Nov 2026

    2.8T total parameters, 104B active; above Claude Opus 5.5 on 16 of the 37 non-internal rows of Moonshot’s own table lab-reported. Our story

    $2.00 / $10.00 per 1M tokens
  • Thinking Machines
    Inkling 2 Mini
    Released 20 Nov 2026

    48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified (bash only), leading open-weights models under 100B total parameters lab-reported. 26.4B parameters, about 20GB, Apache 2.0. Our story

    Open weights
  • Tencent
    Hy4
    Released 4 Nov 2026

    75.2 on APEX-Agents v1.1 against 74.8 for Claude Opus 5.5, in Tencent’s testing lab-reported; Tencent itself says “a notable gap remains” to top proprietary models. Our story

    $0.834 / $2.501 per 1M tokens
  • Ricursive Intelligence
    Ripple-65B Preview
    Weights released 4 Nov 2026

    80.5 on Terminal-Bench 2.1 against 80.1 for the 501-billion-parameter Reflection Beam; top score in 3 of the 22 rows of its own table lab-reported. Apache 2.0. Our story

    Open weights

Price per million tokens

List prices per million input and output tokens, standard tier, short prompts, no caching or batch discounts. Dated by the page we read; “as reported” rows come from the lab’s own post via our newsroom.

ModelInputOutputAs of
Claude Fable 5.2Anthropic$8$4021 Nov, Anthropic pricing page
GPT-6.2 AstraOpenAI$8$4021 Nov, OpenAI pricing page
GPT-6 AstraOpenAI$10$5021 Nov, OpenAI pricing page
Claude Opus 5.5Anthropic$4$2021 Nov, Anthropic pricing page
Gemini 4 NeonGoogle$3$1521 Nov, Gemini API pricing page
Gemini 4 CarbonGoogle$2 ($4 from 1 Dec)$10 ($20 from 1 Dec)21 Nov, Gemini API pricing page
Claude Sonnet 5.5Anthropic$2$1021 Nov, Anthropic pricing page
GPT-6.2 SolOpenAI$2$1021 Nov, OpenAI pricing page
Kimi K3.1 ProMoonshot$2.00$10.0023 Oct, Moonshot post on X, as reported
Gemini 4 HydrogenGoogle$1.50$7.5021 Nov, Gemini API pricing page
Grok 4.7xAI$2$621 Nov, xAI models page
Qwen4-MaxAlibaba$1.5$4.521 Nov, Model Studio prices (International)
DeepSeek V4 ProDeepSeek$1.32 peak / $0.66 off-peak$3.96 / $1.9821 Nov, DeepSeek pricing page (cache miss)
Tencent Hy4Tencent$0.834$2.5014 Nov, Tencent release page, as reported
Mistral Large 4Mistral$0.7$2.121 Nov, Mistral pricing page
DeepSeek V4.1 FlashDeepSeek$0.3 peak / $0.15 off-peak$1.2 / $0.621 Nov, DeepSeek pricing page
Inkling 2 SmallThinking Machines$0.30$1.006 Nov, per Artificial Analysis
Claude Haiku 5.5Anthropic$0.10$0.5021 Nov, prompts up to 100,000 tokens
Muse EmberMeta$0.10$0.509 Nov, Meta Model API, as reported

Sources: Anthropic, OpenAI, Google, xAI, Alibaba Cloud, DeepSeek and Mistral pricing pages, read 21 November 2026. DeepSeek’s off-peak rates are half of peak. Anthropic says its newer tokenizer produces approximately 30% more tokens for the same text, so the same job can cost different amounts at the same list price. Newsroom guide: AI token prices, model by model.

Independent benchmarks: what third parties measure

Scores the labs did not run themselves. Each is a different test with a different method, so the leader changes from one to the next.

Arena: Agent Arena Index

Built on over 2.9 million user sessions. On Arena’s leaderboard dated 18 November, Claude Fable 5.2 (Max) is first at 1,448 points, 112 ahead of Claude Opus 5.5 (Max) at 1,336. In its 20 November reading Arena ranked Gemini 4 Neon second with 1,336 points at a median $2.82 per task, and Gemini 4 Carbon (High) fourth at 1,303. Arena’s relaunch · Neon at #2

Epoch AI: Epoch Capabilities Index

Epoch Capabilities Index, 17 Nov 2026: GPT-6.2 Astra 174.17, Claude Fable 5.2 173.5, Gemini 4 Carbon 169.78, GPT-6 Astra 169.2, Claude Fable 5.1 167.09, Gemini 4 Argon 166.57
According to Epoch AI, the scores for the newest models are “based on testing we did prior to launch using only our own benchmarks” and “may change in the future”. The 0.67-point gap sits inside each model’s confidence interval. Our story

ARC Prize: ARC-AGI-3

ARC-AGI-3 Semi-Private score, Standard vs Provider Adapter harness: GPT-6.2 Astra 70.31% to 100%, Claude Fable 5.2 81.17% to 97.72%, Gemini 4 Carbon 65.12% to 99.09%, GPT-6 Astra 62.71% to 99.95%
ARC Prize verified GPT-6.2 Astra as the first model to 100% on ARC-AGI-3, with its Provider Adapter harness, in a $4.1K run; on the Standard harness, which ARC Prize says gives “a consistent, apples-to-apples comparison across providers”, Claude Fable 5.2 leads at 81.17% against 70.31%. Our story

Artificial Analysis: Intelligence Index

Artificial Analysis Intelligence Index, 11 Nov 2026, cost per index task: Claude Fable 5.2 max (65) $26.54, Fable 5.2 high (62) $9.08, Gemini 4 Carbon high (60) $7.76, GPT-6 Astra max (57) $6.24, GPT-6.1 Sol max (55) $1.27
The index (version 5.0.1) combines nine evaluations, including Terminal-Bench 4.0 and HLE-Diamond. On 18 November GPT-6.2 Astra scored 62, 95% of Fable 5.2’s top score at 15% of its cost per task, $4.00 against $26.54. 11 Nov leaderboard · GPT-6.2 Astra

SemiAnalysis: GDPx

An agent benchmark of 206 tasks across 11 domains, built with 14 data partners. At launch on 2 November, Claude Opus 5.5 led at 12.0% (“even the best model scores just 12%”, SemiAnalysis wrote) and Gemini 4 Carbon was second at 10.9%, at $16.29 a task against $51.41. On 10 November SemiAnalysis said Claude Fable 5.2 debuted at the top at 16.4%. GDPx launch · Fable 5.2 on GDPx

METR: how long a task an agent can finish

METR says the length of task frontier agents complete with 50% success is doubling every 3.61 months, or 110 days (95% confidence interval 2.08 to 4.94 months), reaching up to 27 hours on its verifiable tasks; it tested 15 models. On tasks without clear feedback, “AI agents appear to have worse judgment than human experts”. Our story

Security tests

Vals AI scored DeepSeek V4.1 Flash 1017 at 16.3 of 21 objectives on its offensive-cyber task, against 15.7 for GPT-6 Sol Daybreak (20 October). Epoch AI said open-weight GLM 5.3 built working attacks on six Firefox releases at a marginal cost of about $1,000 each at API pricing (29 October). Vals · Epoch on GLM 5.3

Local and open models: speed and cost claims

Open weights mean you can download the model and run it on your own hardware instead of renting it by the token. Speed claims here are attributed to whoever measured them.

Inkling 2 Mini decode speed as reported by EXO Labs: 210 tokens per second on a 96GB M5 Ultra Mac Studio (prefill 3,620), 125 on a 48GB M5 Max MacBook Pro (prefill 1,750)
According to EXO Labs (post on X, 21 November 2026), working with Thinking Machines. At 210 tokens a second a 1,000-token answer takes about 4.8 seconds; at 125, about 8.0 seconds. Hugging Face co-founder Julien Chaumond quoted EXO’s post: “And this, my friends, is how local and open source AI is going to win”. Our story · EXO Labs’ post
  • The Claude comparison: EXO says Inkling 2 Mini built a native weather app 1.8x faster than Claude Sonnet 5.5 on an M5 Ultra Mac Studio and 11x cheaper at API pricing. That is EXO’s own test: it has not published the task, the number of runs, how it timed “faster”, or which API prices it used.
  • The model: Thinking Machines reports 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified for Inkling 2 Mini, a 26.4B-parameter looped model of about 20GB that its card lists for a DGX Spark or Apple silicon with 32GB or more. Its larger Inkling 2 Small (594B total, 17B active) is on Hugging Face under Apache 2.0 and lists at $0.30 and $1.00 per million tokens per Artificial Analysis. Our story
  • Serving speed in the cloud: Baseten says its open-source engine decodes DeepSeek-V4.1-Flash at about 2,060 tokens a second on eight B300 GPUs and 1,740 on four GB300s, for a single request, and prefills 1,000,000 tokens in 11.7 seconds. Our story
  • A trillion parameters on a Mac: Unsloth says Mistral Large 4 runs on a Mac Studio with its 1-bit GGUF, which it says retains 91% top-1 accuracy. Our story
  • Licences differ: Inkling 2 Mini, Inkling 2 Small and Ripple-65B are Apache 2.0. Meta (Muse Spark), Moonshot (Kimi K3.1 Pro) and MiniMax (M3.1 Flash, 428B total, 23B active) require a separate agreement for large commercial API hosts, and Mistral for Model-as-a-Service businesses.

What benchmarks do and don’t tell a business buyer

  • Price per task beats price per token. On the Artificial Analysis index, Fable 5.2 at its high setting scores 62 for $9.08 a task; its max setting scores 65 for $26.54. The last 3 points cost $17.46 more per task. Arena puts Gemini 4 Neon’s median agent task at $2.82.
  • The harness matters as much as the model. The same GPT-6.2 Astra scores 70.31% or 100% on ARC-AGI-3 depending on ARC Prize’s harness. Google says Gemini 4 Neon goes from 12.8% to 39.0% on SWE-sweep inside its own Antigravity tool. Test the model in the setup you will actually use.
  • Tokens are not equal across labs. Anthropic says its newer tokenizer produces approximately 30% more tokens for the same text; OpenAI says GPT-6.2 Astra uses fewer output tokens. Same list price, different bill.
  • Lab tables are chosen by the lab. Kimi K3.1 Pro beats Opus 5.5 on 16 of 37 rows of Moonshot’s table; Qwen4-Max tops 4 of 16 panels in Alibaba’s. Look for the rows they lose.
  • Contamination. Models may have seen the test material in training. On its own MedChemBench, OpenAI says “Prior exposure to the published sources cannot be ruled out”, and METR warns that easy-to-run evals “could overestimate real-world capabilities”.
  • Speed claims need a method. Tokens a second depend on the hardware, the quantisation, the prompt length and whether one or many requests run at once. Baseten’s 2,060 is a single-request number; EXO’s weather-app comparison has no published method.

What to watch

  1. 1 December: Gemini 4 Carbon and Gemini 4 Argon move from $2 and $10 to $4 and $20 per million tokens on Google’s pricing page. At the new rate Carbon matches Claude Opus 5.5 per token.
  2. 30 November: OpenAI’s pricing page says GPT-5.6 Sol’s promotional pricing ends.
  3. More Gemini 4 models. A Google DeepMind staffer posted “Shipping will continue” on 19 November; Google has not named the next model or its price.
  4. Independent tests of the open models. Inkling 2 Mini’s scores are Thinking Machines’ own and its local speeds are EXO Labs’; independent results were not out on 21 November.
  5. METR’s curve. METR notes that naive extrapolation would reach a 50% time horizon of 200 hours in July 2027, and says that is an extrapolation, not a forecast.

Dates from Google’s and OpenAI’s pricing pages (read 21 November 2026), METR and our newsroom. We update this page when a lab launches a model or changes a price.

AI model benchmarks FAQ

What is the best AI model right now?

It depends on the test, and the top two are close. Anthropic’s Claude Fable 5.2 leads Arena’s Agent Arena Index at 1,448 points (18 November 2026) and the Artificial Analysis Intelligence Index at 65 (11 November). OpenAI’s GPT-6.2 Astra scores 174.17 on Epoch AI’s capability index against 173.5 for Fable 5.2, a gap inside each model’s confidence interval (17 November), and it was the first model to reach 100% on ARC-AGI-3, with the harness that keeps its reasoning state. On ARC Prize’s Standard harness Fable 5.2 leads, 81.17% to 70.31%. Both list at $8 per million input tokens and $40 per million output tokens.

How much does Claude Fable 5.2 cost compared with GPT-6.2 Astra?

The same, per token. As of 21 November 2026 Anthropic’s pricing page lists Claude Fable 5.2 at $8 per million input tokens and $40 per million output tokens, and OpenAI’s lists GPT-6.2 Astra at $8 and $40 (standard tier, short context). Claude Opus 5.5 is $4 and $20; Claude Sonnet 5.5 and GPT-6.2 Sol are both $2 and $10. Per-token prices are not cost per task: Anthropic says its newer tokenizer produces approximately 30% more tokens for the same text, and OpenAI says GPT-6.2 Astra uses fewer output tokens.

What do the Gemini 4 models cost?

On Google’s Gemini API pricing page, read 21 November 2026: Gemini 4 Carbon $2 input and $10 output per million tokens through 30 November 2026, then $4 and $20 from 1 December; Gemini 4 Neon $3 and $15; Gemini 4 Hydrogen $1.50 and $7.50, with a free tier. Arena puts Gemini 4 Neon’s median cost at $2.82 per agent task.

Which frontier model is cheapest?

Among the flagships on this page, by list price per million output tokens on each lab’s own page (21 November 2026): Mistral Large 4 at $2.1, DeepSeek V4 Pro at $3.96 at peak hours ($1.98 off-peak), Qwen4-Max at $4.5 and Grok 4.7 at $6, against $40 for Claude Fable 5.2 and GPT-6.2 Astra. Cheaper per token does not always mean cheaper per finished task.

What is SemiAnalysis’s GDPx benchmark?

GDPx is an agent benchmark SemiAnalysis released on 2 November 2026: 206 tasks across 11 domains, built with 14 data partners, from chip design to finance and biology. At launch Claude Opus 5.5 led at 12.0% and Gemini 4 Carbon was second at 10.9%, at $16.29 a task against $51.41 for Opus 5.5. On 10 November SemiAnalysis said Claude Fable 5.2 debuted at the top at 16.4%.

What is Arena’s Agent Arena Index?

Arena’s agent leaderboard, relaunched in November 2026 and built on over 2.9 million user sessions. It combines task completion, praise versus complaint, steerability and bash reliability. On its 18 November leaderboard Claude Fable 5.2 (Max) led at 1,448 points; by 20 November Arena ranked Gemini 4 Neon second at 1,336, at a median $2.82 per task.

Can you run a frontier-class AI model on a laptop?

Small open models now run fast on high-end Macs. According to EXO Labs, Thinking Machines’ Inkling 2 Mini runs at 210 tokens a second (decode) and 3,620 tokens a second (prefill) on a 96GB M5 Ultra Mac Studio, and at 125 and 1,750 on a 48GB M5 Max MacBook Pro (post on X, 21 November 2026). EXO also says the model built a native weather app 1.8x faster than Claude Sonnet 5.5 and 11x cheaper at API pricing; that is EXO’s own test and it has not published the method. Thinking Machines reports 48.2% on DeepSWE v1.1 for the roughly 20GB model.

Are lab-reported AI benchmarks reliable?

Treat them as claims. Labs choose which tests to show: in Moonshot’s own table Kimi K3.1 Pro beats Claude Opus 5.5 on 16 of 37 rows, and Qwen4-Max tops 4 of the 16 panels in Alibaba’s chart. Settings matter: Thinking Machines notes that “evaluation setups vary by model”. And training data can overlap with the test: on its own MedChemBench, OpenAI says “Prior exposure to the published sources cannot be ruled out”. Independent leaderboards (Arena, Artificial Analysis, Epoch AI, ARC Prize, SemiAnalysis, METR) are a better guide, and even they disagree.

How fast are AI agents improving?

METR said on 20 November 2026 that the length of task frontier agents can finish with 50% success is doubling every 3.61 months (110 days; 95% confidence interval 2.08 to 4.94 months), reaching up to 27 hours on its verifiable tasks. It also found that on tasks without clear feedback “AI agents appear to have worse judgment than human experts”, and warns its easy-to-run evals “could overestimate real-world capabilities”.

All our model launch and benchmark stories

Every piece our newsroom has published on model launches, benchmarks, prices and speed since 1 October, by theme, newest first. New launch pieces link back here.

Model launches and list prices

Independent benchmarks and leaderboards

Related hubs and trackers

Sources and method. Prices are read from each lab’s own pricing page on 21 November 2026 (Anthropic, OpenAI, Google, xAI, Alibaba Cloud, DeepSeek, Mistral) or, where marked, from the lab’s own post as reported by our newsroom. Lab-reported scores come from the lab’s launch post or model card (Anthropic, OpenAI, Google, Alibaba, Mistral, Moonshot, Thinking Machines, Tencent, Ricursive Intelligence), via the linked newsroom piece. Third-party scores are attributed to the body that measured them: Arena, Artificial Analysis, Epoch AI, ARC Prize, SemiAnalysis, METR, Vals AI. Speed claims are attributed to EXO Labs, Baseten and Unsloth. Figures marked as ours are our arithmetic. Every figure on this page has a trace row with its source, verbatim quote and date.

Scroll to Top
FourWeekMBA