LLM Leaderboard 2026: Every Frontier Model’s Own Scores, Prices and Context Windows, Sourced

LLM Leaderboard 2026

Last updated: October 11, 2026, 13:00 UTC | Update cadence: within 48 hours of any qualifying release, plus a verification sweep every Monday and Thursday, plus a full monthly re-verification | Maintained by: Axis Intelligence Research & Sarah Mitchell, AI/ML Editor — 14 models · 24 cross-source observations · as of 2026-10-11 | CC BY 4.0

---

Quick answer

As of October 11, 2026, Claude Sonnet 5.5 posts the highest developer-reported Terminal-Bench 4.0 score among 14 tracked frontier models, at 70.6%, according to Anthropic. Axis Intelligence Research finds that developers publish only 46.4% of the scores on its four-benchmark core set, and that the same model’s score differs between labs’ documents in 5 of 9 cases checked.

Key findings

1. Claude Sonnet 5.5 leads Terminal-Bench 4.0 at 70.6%, ahead of Claude Opus 5.5 at 66.4%, according to Anthropic’s September 2026 announcements.
2. Developers publish 26 of the 56 core-set scores, a Self-Report Coverage of 46.4%, Axis Intelligence Research finds across 14 frontier models as of October 11, 2026.
3. GPT-6 Astra is the only tracked model whose developer reports all four core benchmarks; OpenAI publishes none of the four for GPT-6.1 Sol or GPT-6 Luna.
4. The same score is not the same number everywhere: Claude Opus 5.5’s Terminal-Bench Science 0.1 result reads 58.7% in Anthropic’s announcement and 63.3% in Google’s Gemini 4 Argon table, a 4.6-point Cross-Source Divergence.
5. List prices span a 100-fold range: the Axis Blended Price runs from $0.20 per million tokens (Claude Haiku 5.5, GPT-6 Luna) to $20.00 (Claude Fable 5.1, GPT-6 Astra).

Which LLM Is the Best Right Now?

There is no single answer, and the table above is built to stop you from pretending there is one. Sort it by the column that matches your work.

On agentic terminal work, the only benchmark that most frontier labs now report about their own models, the developer-reported order is Claude Sonnet 5.5 (70.6%), Claude Opus 5.5 (66.4%), GPT-6 Astra (57.9%), Gemini 4 Argon (57.4%) and Claude Fable 5.1 (55.8%). Sonnet 5.5 finishing above Opus 5.5 is Anthropic’s own number, printed in its own Sonnet 5.5 announcement, and the same page cautions that Opus 5.5 remains stronger on open-ended work that needs sustained judgment. Read the score as a measurement of one harness, not a verdict on the model.

On Humanity’s Last Exam with tools, Claude Opus 5.5 reports the top figure at 67.7%, followed by Claude Fable 5.1 at 65.6% and Claude Sonnet 5.5 at 64.5%. DeepSeek-V4.1-Flash, an open-weight model, reports 63.9% in DeepSeek’s September 10 change log — within 3.8 points of the top proprietary result on this column.

Terminal-Bench 4.0 tells the opposite story for open weights. DeepSeek-V4.1-Flash reports 31.2% there, 39.4 points below Sonnet 5.5. A model that is close on a reasoning exam and far behind on long terminal tasks is a familiar pattern: the exam rewards knowledge, the terminal rewards hours of not losing the thread.

Why Axis does not publish a single overall score

Each lab runs its own evaluations, with its own harness, its own effort setting and, in several cases, its own safety fallbacks switched on. Anthropic notes that Opus 5.5 was evaluated with production safeguards that rerouted some tasks to other models; OpenAI reports “the maximum at any effort”; Google uses its highest thinking setting. Averaging those numbers into one index would merge incompatible methodologies and produce a figure that means nothing. Axis publishes the columns side by side and leaves the weighting to the reader whose workload defines it.

How Many Benchmarks Do AI Labs Actually Report?

Fewer than you would think, and not the same ones. Axis Intelligence Research tracks four core benchmarks — Terminal-Bench 4.0, Humanity’s Last Exam with tools, FrontierCode 1.1 Main and GPQA Diamond — and records a score only when the model’s own developer prints it.

Core benchmark	Models whose developer reports it	Coverage	Source
Terminal-Bench 4.0	10 of 14	71.4%	Developer announcements and model cards (CSV source_url)
Humanity’s Last Exam, with tools	8 of 14	57.1%	Same
FrontierCode 1.1 Main	5 of 14	35.7%	Same
GPQA Diamond	3 of 14	21.4%	Same
All four combined (Self-Report Coverage)	26 of 56	46.4%	Axis calculation

GPQA Diamond, for years the default science benchmark in launch posts, now appears for only three tracked models: GPT-6 Astra (96.0%), Kimi K3 (93.5%) and DeepSeek-V4.1-Flash (90.9%). None of the Anthropic, Google, SpaceXAI or Mistral AI documents Axis reviewed for the tracked models prints it. The benchmark that used to be table stakes has quietly left the table.

At the other end sit GPT-6.1 Sol and GPT-6 Luna, with zero core-set numbers. OpenAI’s announcements for both models present results as charts and relative claims — “matches GPT-6 Astra at roughly one-fifth of the cost” — rather than figures a reader can copy into a spreadsheet. That is a disclosure choice, and the leaderboard records it as one.

This is the baseline reading of Self-Report Coverage. It has no history yet; the next reading publishes with the November monthly review.

Do AI Labs Report the Same Scores for the Same Model?

Not always. When one lab prints a rival’s score in its comparison table, that figure can differ from what the rival printed about itself. Axis Intelligence Research logged every case where at least two labs’ documents give a score for the same tracked model on the same benchmark.

Model · benchmark	Reported values (by whom)	Divergence
Claude Opus 5.5 · Terminal-Bench Science 0.1	58.7% Anthropic · 63.3% Google	4.6 pts
GPT-6 Astra · Terminal-Bench Science 0.1	64.6% OpenAI (Astra post), Anthropic · 68.1% OpenAI (GPT-6.1 Sol post), Google	3.5 pts
Claude Fable 5.1 · Terminal-Bench 4.0	55.8% Anthropic, OpenAI · 57.9% Google, SpaceXAI	2.1 pts
Claude Fable 5.1 · HLE with tools	65.6% Anthropic · 65.0% OpenAI	0.6 pts
GPT-6 Astra · Terminal-Bench 4.0	57.9% OpenAI, Anthropic · 58.2% Google	0.3 pts
Claude Opus 5.5 · Terminal-Bench 4.0	66.4% Anthropic, Google	0
Gemini 3.8 Flash · Terminal-Bench 4.0	19.1% Google, OpenAI	0
GPT-6 Astra · HLE with tools	57.2% OpenAI, Anthropic	0
Claude Fable 5.1 · Terminal-Bench Science 0.1	52.6% Anthropic, OpenAI, Google	0

Five of nine pairs diverge (55.6%); the median divergence is 0.3 points and the maximum 4.6 points. Most gaps have a mundane explanation that the documents themselves disclose. Google’s Gemini 4 Argon evaluation methodology states that its Terminal-Bench figures for non-Gemini models come from the benchmark authors’ public leaderboard, at each model’s best thinking level — not from the vendors’ own posts. Anthropic, for its part, ran Opus 5.5 with production safeguards that handed some tasks to older models, which it says likely lowers the score.

The most telling row is GPT-6 Astra on Terminal-Bench Science: OpenAI itself printed 64.6% on September 3 and 68.1% on September 29. A lab revising its own flagship’s number inside four weeks is normal model maintenance. A reader quoting the first number in November is not. That is why every cell in this leaderboard carries the document it came from.

How Much Do Frontier LLMs Cost Per Million Tokens?

List prices below are the standard rates the developer publishes for its base tier, verified on October 11, 2026. The Axis Blended Price weights input three-to-one over output, which approximates retrieval-heavy and agentic workloads where prompts dwarf replies.

Model	Input / output, list (USD per 1M)	Axis Blended Price	Basis	Source
Claude Fable 5.1	$10 / $50	$20.00	Standard	Anthropic
GPT-6 Astra	$10 / $50	$20.00	Prompts ≤272K tokens	OpenAI
Claude Opus 5.5	$4 / $20	$8.00	Standard	Anthropic
Gemini 4 Argon	$4 / $20	$8.00	Post-introductory; launches at $2 / $10	Google
Kimi K3	$3 / $15	$6.00	Cache-miss input	Moonshot AI
Claude Sonnet 5.5	$2 / $10	$4.00	Standard	Anthropic
GPT-6.1 Sol	$2 / $10	$4.00	Prompts ≤272K tokens	OpenAI
Gemini 3.8 Flash	$1.50 / $7.50	$3.00	Regular; currently $0.75 / $3.75 introductory	Google
Grok 4.7	$2 / $6	$3.00	Standard	SpaceXAI
Mistral Large 4	$1.36 / $4.18	$2.07	Original; currently on sale at $0.68 / $2.09	Mistral AI
DeepSeek-V4-Pro-0813	$1.32 / $3.96	$1.98	Peak; off-peak is half	DeepSeek
DeepSeek-V4.1-Flash	$0.30 / $1.20	$0.53	Peak; off-peak is half	DeepSeek
Claude Haiku 5.5	$0.10 / $0.50	$0.20	Prompts ≤100K tokens	Anthropic
GPT-6 Luna	$0.10 / $0.50	$0.20	Prompts ≤272K tokens	OpenAI

The median Axis Blended Price across the 14 models is $3.50 per million tokens. The pricing that matters most for agents — cache reads — moved twice in the last month alone: Anthropic halved Claude Sonnet 5.5’s cache-read rate to $0.10 per million on October 7, and DeepSeek cut prices with the V4.1-Flash release on September 10. Token price is also not task price. Anthropic and OpenAI both argue in their launch posts that fewer tokens per task matter more than the per-token rate, and neither claim can be checked from a price sheet. For workload math, the LLM API Cost Calculator takes your own token mix.

Which Open-Weight LLMs Are on the Leaderboard?

Three tracked models have downloadable weights today: Kimi K3 (Moonshot AI, under the Kimi K3 License), DeepSeek-V4.1-Flash and DeepSeek-V4-Pro-0813 (both with weights repositories on Hugging Face; Axis has not yet verified their license text). Mistral Large 4 is in public preview through Mistral’s API, and Mistral’s announcement promises weights “by the end of the month”; the leaderboard marks it open_announced until the files exist.

Every open-weight model on the board comes from a lab outside the United States. The cheapest of them, DeepSeek-V4.1-Flash, lists at $0.30 input and $1.20 output per million at peak hours.

Which Models Have the Largest Context Windows?

Most frontier models now accept about one million tokens of input. OpenAI’s GPT-6 family lists 1,050,000 tokens, Google’s Gemini 3.8 Flash and Moonshot’s Kimi K3 list 1,048,576, and Anthropic’s four Claude 5.x models list one million. Grok 4.7 lists 500,000. Google has not published a context window for Gemini 4 Argon; it has published an output limit of one million tokens, up from 64K on earlier models.

Output limits vary far more than input limits: 65,536 tokens for Gemini 3.8 Flash, 128,000 for the Claude and GPT-6 families, 384,000 for DeepSeek’s V4-series API. Long input is now standard. Long output is where the models still differ.

Leaderboard entries

Reverse chronological by release date. Each entry: date · headline · summary · amount · primary source. Entries older than 30 days move to the archive block.

October 7, 2026 — Anthropic releases Claude Haiku 5.5

Anthropic’s smallest model posts 39.2% on Terminal-Bench 4.0, 57.4% on Humanity’s Last Exam with tools and 46.4% on FrontierCode 1.1 Main. It is Anthropic’s first Haiku model with an adjustable effort setting. The same announcement cut Claude Sonnet 5.5’s cache-read price in half. Price: $0.10 / $0.50 per 1M (prompts ≤100K tokens); $0.50 / $2.50 above · Source: Anthropic

October 6, 2026 — Mistral AI previews Mistral Large 4

Mistral AI opened a public preview of Mistral Large 4, a mixture-of-experts model its model page lists at 1.05T total and 52B active parameters, with a one-million-token context. Mistral reports 28.3% on Terminal-Bench 4.0. Weights are promised by the end of October. Price: $0.68 / $2.09 per 1M on sale; $1.36 / $4.18 original · Source: Mistral AI

September 30, 2026 — Google announces Gemini 4 Argon

Google’s new flagship reports 57.4% on Terminal-Bench 4.0 and ships with a one-million-token output limit. Access is limited to defenders in Google’s Fairwind Program; paid API customers come later, on a date Google has not given. No context window or API model ID has been published. Price: $2 / $10 per 1M introductory; $4 / $20 afterward · Source: Google

September 29, 2026 — OpenAI releases GPT-6.1 Sol

OpenAI’s mid-tier upgrade lists a 1,050,000-token context window and 128,000 output tokens. Its announcement reports benchmark results as charts and relative claims, so none of the four core benchmarks has a printed figure. Cached input costs $0.10 per million. Price: $2 / $10 per 1M (prompts ≤272K tokens) · Source: OpenAI

September 28, 2026 — Anthropic releases Claude Sonnet 5.5

Sonnet 5.5 reports 70.6% on Terminal-Bench 4.0 — the highest developer-reported figure on the board — plus 64.5% on Humanity’s Last Exam with tools and 52.1% on FrontierCode 1.1 Main at xhigh effort. Anthropic says Opus 5.5 remains stronger on complex open-ended work. Price: $2 / $10 per 1M · Source: Anthropic

September 22, 2026 — Anthropic releases Claude Opus 5.5

Opus 5.5 reports 66.4% on Terminal-Bench 4.0 at xhigh effort, 67.7% on Humanity’s Last Exam with tools and 54.4% on FrontierCode 1.1 Main. Anthropic evaluated it with production safeguards that rerouted some tasks to older models. Input and output prices fell 20% from Opus 5. Price: $4 / $20 per 1M · Source: Anthropic

September 22, 2026 — OpenAI releases GPT-6 Luna

OpenAI’s low-cost GPT-6 model lists a 1,050,000-token context and a May 18, 2026 knowledge cutoff. Its announcement reports DeepSWE v1.1 and OSWorld 2.0 results but none of the four core benchmarks. Price: $0.10 / $0.50 per 1M (prompts ≤272K tokens) · Source: OpenAI

September 21, 2026 — SpaceXAI releases Grok 4.7

Grok 4.7 reports 37.6% on Terminal-Bench 4.0 at xHigh effort and lists a 500,000-token context window. SpaceXAI says it is built on a new, larger base model than Grok 4.6, at the same price. Price: $2 / $6 per 1M · Source: SpaceXAI

September 10, 2026 — DeepSeek releases DeepSeek-V4.1-Flash

DeepSeek reports 31.2% on Terminal-Bench 4.0, 63.9% on Humanity’s Last Exam with tools and 90.9% on GPQA Diamond. Weights are published on Hugging Face. API prices fell with the release. Price: $0.30 / $1.20 per 1M at peak; half off-peak · Source: DeepSeek

September 3, 2026 — OpenAI releases GPT-6 Astra

GPT-6 Astra is the only tracked model with all four core scores from its developer: 57.9% on Terminal-Bench 4.0, 57.2% on Humanity’s Last Exam with tools, 53.3% on FrontierCode 1.1 Main and 96.0% on GPQA Diamond, each the maximum at any effort. Price: $10 / $50 per 1M (prompts ≤272K tokens) · Source: OpenAI

September 2, 2026 — Google releases Gemini 3.8 Flash

Google’s Flash model reports 19.1% on Terminal-Bench 4.0 and lists a 1,048,576-token input limit with 65,536 output tokens. It is sold at an introductory price, half its regular rate. Price: $0.75 / $3.75 per 1M introductory; $1.50 / $7.50 regular · Source: Google

September 1, 2026 — Anthropic releases Claude Fable 5.1

Anthropic’s most capable generally available model reports 55.8% on Terminal-Bench 4.0, 65.6% on Humanity’s Last Exam with tools and 50.3% on FrontierCode 1.1 Main, per the benchmark table Anthropic published with Opus 5.5. Cybersecurity and biology requests can be routed to other models. Price: $10 / $50 per 1M · Source: Anthropic

Archive (older than 30 days)

2 archived entries

August 13, 2026 — DeepSeek ships DeepSeek-V4-Pro-0813

The general-availability release of DeepSeek-V4-Pro reports 60.0% on Humanity’s Last Exam with tools. Weights are published on Hugging Face. It does not report Terminal-Bench 4.0. Price: $1.32 / $3.96 per 1M at peak; half off-peak · Source: DeepSeek

July 2026 — Moonshot AI releases Kimi K3

Moonshot’s 2.8T-parameter open-weight model reports 93.5% on GPQA Diamond and 56.0% on HLE-Full with tools. Weights are published under the Kimi K3 License. The exact release day is pending a dated primary source. Price: $3 / $15 per 1M (cache-miss input) · Source: Moonshot AI

How We Curate This List

What counts. A model enters when (1) its developer has published a primary document — announcement, model card or API documentation — that Axis fetched and read; (2) it is the developer’s current model at a flagship, mid or efficiency tier; and (3) it can be used by outside customers today, through general availability, a public preview or a limited-access program.

What does not count. Superseded tiers are retired: GPT-6 Sol left the board when OpenAI described GPT-6.1 Sol as its upgrade. Image, video, audio and embedding models are out of scope. Specialist variants with separate access tiers — cyber-specific or life-sciences models — are out of scope. Release events themselves belong to the AI Model Release Tracker; this leaderboard holds each model’s current state, and no row duplicates a release-tracker entry.

What a score is. A benchmark cell holds the figure the model’s own developer printed for that model, with the effort setting the developer names. A rival’s figure for the same model is never used in the table; it is logged separately as a divergence observation. Scores produced and published by third-party aggregator leaderboards are excluded. Benchmark variants are not merged: HLE-Verified is not Humanity’s Last Exam, OSWorld 2.0 is not OSWorld 2.1, Terminal-Bench 2.1 is not 4.0.

What a price is. The developer’s standard list rate for the base tier, with any promotional or introductory rate shown alongside and the threshold or time-of-day condition stated.

What Axis does not do. Axis does not run these evaluations. Every score here is self-reported. That is the honest basis of this page, and it is exactly why the Cross-Source Divergence log exists.

Methodology

Self-Report Coverage. SRC = (number of core-benchmark scores the developers publish for the tracked models) ÷ (number of models × 4). Core set: Terminal-Bench 4.0, Humanity’s Last Exam with tools, FrontierCode 1.1 Main, GPQA Diamond. Baseline reading: 26 ÷ (14 × 4) = 46.4%. Chosen because these four are the benchmarks most often printed across the tracked labs’ current documents.

Cross-Source Divergence. For each model–benchmark pair reported in at least two organisations’ primary documents, CSD = highest reported value − lowest reported value, in percentage points. Baseline: 9 pairs, 5 non-zero, median 0.3 points, maximum 4.6 points. Pairs are drawn from Terminal-Bench 4.0, HLE with tools and Terminal-Bench Science 0.1, the benchmarks on which labs most often print each other’s models.

Axis Blended Price. ABP = (3 × input price + output price) ÷ 4, in USD per million tokens, at the developer’s list price for the base tier. Introductory, sale and off-peak rates are recorded but not used in ABP, so a promotion ending does not move the metric.

Limitations. Developer-reported scores use different harnesses, effort settings and safeguard configurations; a shared column name does not guarantee a shared test. Coverage measures disclosure, not capability. ABP ignores cache pricing, which dominates agentic workloads. The model set is selected by tier, not by performance.

About this dataset

llm-leaderboard.csv — 38 rows: 14 model rows (one per leaderboard entry, with specs, prices, core scores and source URLs) and 24 divergence_observation rows (one per organisation-reported score in the Cross-Source Divergence log). Every row carries source_url, retrieved_date and is_primary. Licensed CC BY 4.0. Cite as: Axis Intelligence Research, LLM Leaderboard 2026, 2026.

Review trigger: any frontier release with a primary document, any list-price change on a tracked model, any developer revising a printed score, or the release of Mistral Large 4 weights.

Frequently asked questions

What is the best LLM in October 2026?

On developer-reported Terminal-Bench 4.0, Claude Sonnet 5.5 leads at 70.6%. On Humanity’s Last Exam with tools, Claude Opus 5.5 leads at 67.7%. On GPQA Diamond, GPT-6 Astra leads at 96.0% among the three models whose developers report it.

Which LLM is the cheapest?

Claude Haiku 5.5 and GPT-6 Luna both list at $0.10 per million input tokens and $0.50 per million output tokens, an Axis Blended Price of $0.20.

What is the best open-weight LLM?

On the core benchmarks, DeepSeek-V4.1-Flash reports the highest open-weight scores on the board: 63.9% on HLE with tools and 31.2% on Terminal-Bench 4.0. Kimi K3 reports 93.5% on GPQA Diamond.

Are LLM leaderboard scores reliable?

They are what the developers published, under conditions the developers chose. Axis found that the same model’s score differs between labs’ documents in 5 of 9 cases checked, by up to 4.6 points.

How often is this leaderboard updated?

Within 48 hours of any qualifying release, plus a verification sweep every Monday and Thursday and a full monthly re-verification. Across the 36 days from September 1 to October 7, 2026, tracked labs released a qualifying model every 3.27 days on average — so a weekly cycle would leave the board stale for half of each week.

Does Axis test the models itself?

No. Axis compiles developer-published figures from primary documents, computes its own disclosure, divergence and price metrics, and links every number to its source.