Free tool · Updated 2 Oct 2026

AI Model Leaderboard 2026

Compare 72 AI models on benchmark scores, context windows, prices and speed. Every number links to its source. Anything we could not confirm shows a dash, never a zero.

Models tracked

Providers

With sourced benchmark scores

Open-weight models

Model rankings

Top models by publicly reported score on the benchmark you pick. Hover a bar for its source; click it for the model page.

Terminal-Bench 4.0 · Long jobs in a real terminal

Top 10 of 10 with a score · Self-reported = the lab’s own number

100

70.6%66.4%60.9%57.9%55.8%52.3%42%37.6%19.1%10.3%

Claude Sonnet 5.5

Claude Opus 5.5

Claude Mythos 5.1

GPT-6 Astra

Claude Fable 5.1

Claude Opus 5

Claude Fable 5

Grok 4.7

Gemini 3.8 Flash

Claude Sonnet 5

Top models per task

Best in class for each kind of work, ranked by one public benchmark per task. Models with no verified score are left out, not ranked last.

Coding

Terminal-Bench 4.0

Reasoning

GPQA Diamond

1 GPT-6 Astra96%
2 GPT-5.6 Sol94.6%
3 Gemini 3.1 Pro94.3%
4 GPT-5.593.6%
5 Kimi K393.5%
6 GPT-5.6 Terra92.9%

Agents

OSWorld 2.0

Vision

MMMU-Pro

Long context

Context window, tokens

1 Llama 4 Scout10M
2 GPT-6 Astra1.05M
3 GPT-6.1 Sol1.05M
4 GPT-6 Sol1.05M
5 GPT-6 Luna1.05M
6 GPT-5.6 Sol1.05M

Budget

Blended price per 1M tokens (3 in : 1 out), lower is better

1 GPT-6 Luna$0.20
2 Qwen3.8-Flash$0.23
3 Tencent Hy3$0.231
4 Mistral Small 4$0.2625
5 Llama 4 Scout$0.2925
6 Llama 4 Maverick$0.4225

Benchmark leaderboards

The top models on each benchmark. Click a row for the full model page.

DeepSWE v1.1

17 scored

GPQA Diamond

27 scored

Humanity’s Last Exam (no tools)

23 scored

Terminal-Bench 2.1

17 scored

Humanity’s Last Exam (with tools)

16 scored

SWE-bench Pro

16 scored

SWE-bench Verified

12 scored

Terminal-Bench 4.0

10 scored

OSWorld-Verified

8 scored

Fastest and most affordable

Median output speed measured by Artificial Analysis on each provider’s own API; prices from each provider’s pricing page.

Fastest, tokens per second

Model	Provider	Tokens/s	In $/1M
Gemini 3.5 Flash-Lite	Google	360	$0.30
Gemini 3.7 Flash	Google	299	$0.75
Gemini 2.5 Flash-Lite	Google	297	$0.10
Gemini 3.1 Flash-Lite	Google	280	$0.25
Gemini 3.8 Flash	Google	249	$0.75
GPT-5.4 mini	OpenAI	213	$0.75
Command A+	Cohere	211	–

Cheapest input price

Model	Provider	In $/1M	Out $/1M
GPT-6 Luna	OpenAI	$0.10	$0.50
Tencent Hy3	Tencent	$0.132	$0.528
Qwen3.8-Flash	Alibaba	$0.15	$0.47
Mistral Small 4	Mistral	$0.15	$0.60
Llama 4 Scout	Meta	$0.17	$0.66
GPT-5.6 Luna	OpenAI	$0.20	$1.20
Llama 4 Maverick	Meta	$0.24	$0.97

Model comparison

Every model we track. Sort by any column, filter by provider, licence or task, and tick up to four to compare side by side.

72 of 72 models. Tick up to 4 to compare.

		Licence						Best for
GPT-6.1 Sol	OpenAI	Proprietary	1.05M	$2.00	$10.00	64	Sep 2026	CodingAgents
Claude Sonnet 5.5	Anthropic	Proprietary	1M	$2.00	$10.00	139	Sep 2026	CodingAgents
Claude Opus 5.5	Anthropic	Proprietary	1M	$4.00	$20.00	93	Sep 2026	CodingAgents
GPT-6 Sol	OpenAI	Proprietary	1.05M	$2.00	$10.00	92	Sep 2026	CodingAgents
GPT-6 Luna	OpenAI	Proprietary	1.05M	$0.10	$0.50	131	Sep 2026	BudgetAgents
Grok 4.7	xAI	Proprietary	500K	$2.00	$6.00	79	Sep 2026	CodingAgents
Sakana Fugu Ultra v2	Sakana AI	Proprietary	–	$5.00	$30.00	–	Sep 2026	ReasoningCoding
DeepSeek V4.1 Flash	DeepSeek	Open weight	1.05M	$0.30	$1.20	209	Sep 2026	CodingAgents
GPT-6 Astra	OpenAI	Proprietary	1.05M	$10.00	$50.00	51	Sep 2026	CodingAgents
Gemini 3.8 FlashFree tier	Google	Proprietary	1.05M	$0.75	$3.75	249	Sep 2026	CodingAgents
Muse Spark 1.3	Meta	Proprietary	1.05M	$1.25	$4.25	152	Sep 2026	AgentsCoding
Claude Fable 5.1	Anthropic	Proprietary	1M	$10.00	$50.00	68	Sep 2026	ReasoningAgents
Claude Mythos 5.1Invite only	Anthropic	Proprietary	1M	$10.00	$50.00	–	Sep 2026	Reasoningresearch
Qwen3.8-Flash	Alibaba	Open weight	1M	$0.15	$0.47	–	Aug 2026	BudgetCoding
GLM-5.3	Z.ai	Open weight	1.05M	$1.40	$4.40	70	Aug 2026	CodingAgents
Qwen3.8-27B	Alibaba	Open weight	1M	$0.50	$3.00	46	Aug 2026	CodingAgents
Gemini 3.7 FlashFree tier	Google	Proprietary	1.05M	$0.75	$3.75	299	Aug 2026	CodingAgents
Qwen3.8-Max	Alibaba	Open weight	1M	$2.00	$6.00	39	Aug 2026	CodingAgents
Tencent Hy4 previewPreview	Tencent	Open weight	1.05M	$0.834	$2.50	–	Aug 2026	CodingAgents
Claude Opus 5	Anthropic	Proprietary	1M	$5.00	$25.00	55	Jul 2026	CodingAgents
Gemini 3.6 FlashFree tier	Google	Proprietary	1.05M	$0.75	$3.75	185	Jul 2026	CodingAgents
Gemini 3.5 Flash-LiteFree tier	Google	Proprietary	1.05M	$0.30	$2.50	360	Jul 2026	BudgetAgents
GPT-5.6 Sol	OpenAI	Proprietary	1.05M	$4.00	$20.00	79	Jul 2026	CodingAgents
GPT-5.6 Terra	OpenAI	Proprietary	1.05M	$2.00	$12.00	97	Jul 2026	CodingBudget
GPT-5.6 Luna	OpenAI	Proprietary	1.05M	$0.20	$1.20	124	Jul 2026	Budget

Context window, cost and speed

Full specs side by side, with the pricing details a rate card hides: cache prices, long-prompt tiers and promotions that end.

Model	Context	Max output	In $/1M	Out $/1M	Cached in	Tok/s	Pricing note
Llama 4 ScoutMeta	10M	–	$0.17	$0.66	–	108	Open weightMeta does not sell Llama through its own API. Prices are from Amazon Bedrock (us-east-1, on-demand, model ID meta.llama4-scout-17b-instruct-v1:0); Bedrock batch is $0.085 in / $0.33 out. Bedrock caps output at 8K tokens. Groq retired Llama 4 Scout on 17 Jul 2026 (free/developer tiers) and Together no longer lists it on serverless.
GPT-6 AstraOpenAI	1.05M	128K	$10.00	$50.00	$1.00	51	ProprietaryCache writes are 1.25x input; cache reads 0.1x; 30-minute minimum cache life. Batch and Flex: $5 in / $25 out. Fast mode (formerly Priority): 2x, $20 / $100. Ultrafast (Responses API, service_tier "ultrafast", Astra only): $60 in / $6 cached / $75 cache write / $300 out. Data residency endpoints +10%.
GPT-6.1 SolOpenAI	1.05M	128K	$2.00	$10.00	$0.10	64	ProprietaryCache reads are 0.05x input (95% off), not the usual 0.1x. Cache writes 1.25x input. Batch and Flex: $1 / $5. Fast mode: $4 / $20. OpenAI says an Ultrafast option is coming 'in the coming days'; it is not on the pricing page yet. Data residency +10%.
GPT-6 SolOpenAI	1.05M	128K	$2.00	$10.00	$0.20	92	ProprietarySuperseded by GPT-6.1 Sol, which costs the same but has cheaper cached input ($0.10). Batch and Flex: $1 / $5. Fast mode: $4 / $20. Cache writes 1.25x input; 30-minute minimum cache life. Data residency +10%. Launched at 50% below GPT-5.6 Sol's promotional price.
GPT-6 LunaOpenAI	1.05M	128K	$0.10	$0.50	$0.01	131	ProprietaryBatch and Flex: $0.05 / $0.25. Fast mode: $0.20 / $1.00. Cache writes 1.25x input; 30-minute minimum cache life. Data residency +10%. Launched at 50% below GPT-5.6 Luna's price.
GPT-5.6 SolOpenAI	1.05M	128K	$4.00	$20.00	$0.40	79	ProprietaryPromotional price since 21 Aug 2026 (launch price was $5 in / $30 out). OpenAI says it is 'available at least through November 21, 2026'. Batch and Flex: $2 / $10. Fast mode: $8 / $40. Cache writes 1.25x input; 30-minute minimum cache life. The alias gpt-5.6 routes to this model. Data residency +10%.
GPT-5.6 TerraOpenAI	1.05M	128K	$2.00	$12.00	$0.20	97	ProprietaryCut 20% on 30 Jul 2026 (launch price $2.50 in / $15 out). Batch and Flex: $1 / $6. Fast mode: $4 / $24. Cache writes 1.25x input; 30-minute minimum cache life. Data residency +10%.
GPT-5.6 LunaOpenAI	1.05M	128K	$0.20	$1.20	$0.02	124	ProprietaryCut 80% on 30 Jul 2026 (launch price $1 in / $6 out). Batch and Flex: $0.10 / $0.60. Fast mode: $0.40 / $2.40. Cache writes 1.25x input; 30-minute minimum cache life. Data residency +10%.
GPT-5.5OpenAI	1.05M	128K	$5.00	$30.00	$0.50	86	ProprietaryThere is no separate cache-write charge (pre-GPT-5.6 models). Batch and Flex: $2.50 / $15. Fast mode: 2.5x, $12.50 / $75. Snapshot gpt-5.5-2026-04-23. GPT-5.5 Pro (gpt-5.5-pro) costs $30 / $180. Data residency +10%.
GPT-5.4OpenAI	1.05M	128K	$2.50	$15.00	$0.25	88	ProprietaryThere is no separate cache-write charge. Batch and Flex: $1.25 / $7.50. Fast mode: $5 / $30. Snapshot gpt-5.4-2026-03-05. Data residency +10%.
Gemini 3.8 FlashGoogle	1.05M	66K	$0.75	$3.75	$0.075	249	ProprietaryIntroductory price through 31 Dec 2026. From 1 Jan 2027: $1.50 input, $7.50 output, $0.15 cached input. Cache storage $0.50 per 1M tokens per hour, rising to $1.00 in 2027. Google bills no separate cache write. One flat price up to the full 1M context. Flex is also 50% off; Priority is $1.35 / $6.75 ($2.70 / $13.50 from 2027).
Gemini 3.7 FlashGoogle	1.05M	66K	$0.75	$3.75	$0.075	299	ProprietaryIntroductory price through 31 Dec 2026. From 1 Jan 2027: $1.50 input, $7.50 output, $0.15 cached input. Cache storage $0.50 per 1M tokens per hour, rising to $1.00 in 2027. Same price as 3.8 Flash.
Gemini 3.6 FlashGoogle	1.05M	66K	$0.75	$3.75	$0.075	185	ProprietaryLaunched at $1.50 / $7.50. It now has the same introductory price as 3.7 and 3.8 Flash through 31 Dec 2026, then $1.50 / $7.50 from 1 Jan 2027. Cache storage $0.50 per 1M tokens per hour, rising to $1.00 in 2027.
Gemini 3.5 FlashGoogle	1.05M	66K	$1.50	$9.00	$0.15	209	ProprietaryNo promotion. It costs more than the newer 3.6, 3.7 and 3.8 Flash models. Cache storage $1.00 per 1M tokens per hour. Priority is $2.70 / $16.20.
Gemini 3.5 Flash-LiteGoogle	1.05M	66K	$0.30	$2.50	$0.03	360	ProprietaryOne input price for text, image, video and audio. Cache storage $1.00 per 1M tokens per hour. Priority is $0.54 / $4.50.
Gemini 3.1 Flash-LiteGoogle	1.05M	66K	$0.25	$1.50	$0.025	280	ProprietaryAudio input costs $0.50, or $0.05 cached. Cache storage $1.00 per 1M tokens per hour. Scheduled to shut down on 7 May 2027; Google names 3.5 Flash-Lite as the replacement.
Gemini 3.1 ProGoogle	1.05M	66K	$2.00	$12.00	$0.20	114	ProprietaryPreviewStill preview. A separate endpoint, gemini-3.1-pro-preview-customtools, has the same price. Cache storage $4.50 per 1M tokens per hour. Priority is $3.60 / $21.60 up to 200K.
Gemini 3 FlashGoogle	1.05M	66K	$0.50	$3.00	$0.05	182	ProprietaryPreviewAudio input costs $1.00, or $0.10 cached. Cache storage $1.00 per 1M tokens per hour. Google calls it a legacy model; the deprecations page names gemini-3.6-flash as the replacement, with no shutdown date yet.
Gemini 2.5 ProGoogle	1.05M	66K	$1.25	$10.00	$0.125	122	ProprietaryExisting users onlySince 18 Sep 2026 Google has limited 2.5 access to users who actively used these models before. They are not deprecated and have no shutdown date. Cache storage $4.50 per 1M tokens per hour.
Gemini 2.5 FlashGoogle	1.05M	66K	$0.30	$2.50	$0.03	191	ProprietaryExisting users onlyAudio input costs $1.00, or $0.10 cached. Cache storage $1.00 per 1M tokens per hour. Since 18 Sep 2026 access is limited to existing 2.5 users; the model is not deprecated and has no shutdown date.
Gemini 2.5 Flash-LiteGoogle	1.05M	66K	$0.10	$0.40	$0.01	297	ProprietaryExisting users onlyAudio input costs $0.30, or $0.03 cached. Cache storage $1.00 per 1M tokens per hour. The cheapest Gemini text model listed, but since 18 Sep 2026 access is limited to existing 2.5 users. Not deprecated; no shutdown date.
DeepSeek V4 ProDeepSeek	1.05M	393K	$1.32	$3.96	$0.044	104	Open weightPeak price shown. Off-peak is half: $0.66 in, $1.98 out, $0.022 cache hit. Peak hours are 01:00-04:00 and 06:00-10:00 UTC Monday to Friday, excluding Chinese public holidays; weekends are off-peak all day. This peak/off-peak pricing started 16 Aug 2026 16:00 UTC and replaced the flat $0.435/$0.87 price that DeepSeek made permanent on 22 May 2026 (launch list price was $1.74/$3.48).
DeepSeek V4.1 FlashDeepSeek	1.05M	393K	$0.30	$1.20	$0.006	209	Open weightPeak price shown. Off-peak is half: $0.15 in, $0.60 out, $0.003 cache hit. Peak hours 01:00-04:00 and 06:00-10:00 UTC Monday to Friday, excluding Chinese public holidays. Price took effect 10 Sep 2026 04:00 UTC. Legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are routed here and billed at this price.
GLM-5.3Z.ai	1.05M	131K	$1.40	$4.40	$0.26	70	Open weightCached-input storage is free for a limited time.
GLM-5.2Z.ai	1.05M	131K	$1.40	$4.40	$0.26	88	Open weightSame price as GLM-5.3. Cached-input storage is free for a limited time.
Kimi K3Moonshot	1.05M	1.05M	$3.00	$15.00	$0.30	34	Open weightCache write $3.00 (5-minute TTL, default) or $6.00 (1-hour TTL); cache hits $0.30. Prices exclude taxes.
Tencent Hy4 previewTencent	1.05M	64K	$0.834	$2.50	$0.042	–	Open weightPreviewTencent Cloud TokenHub, Singapore region.
Muse Spark 1.3Meta	1.05M	–	$1.25	$4.25	$0.15	152	ProprietaryMeta Model API Standard tier (prompts not used for training). A Contributor tier, muse-spark-1.3-contributor, costs $0.10 in / $0.20 out / $0.002 cached in exchange for Meta training on your prompts and completions. Web search grounding is $2.50 per 1,000 queries.
Claude Fable 5.1Anthropic	1M	128K	$10.00	$50.00	$0.25	68	ProprietaryCache reads are 0.025x base input ($0.25), down from $1 on Fable 5. Uses the newer tokenizer (about 30% more tokens than pre-Opus 4.7 models for the same text). Safety classifiers can refuse; since 24 Sep 2026 pre-output refusals in the bio, frontier_llm and reasoning_extraction categories are billed. Minimum cacheable prompt: 512 tokens.
Claude Mythos 5.1Anthropic	1M	128K	$10.00	$50.00	$0.25	–	ProprietaryInvite onlyInvite only through Project Glasswing; not open to the public. Same prices as Claude Fable 5.1; cache reads are 0.025x base input. Uses the newer tokenizer (about 30% more tokens than pre-Opus 4.7 models). Minimum cacheable prompt: 512 tokens.
Claude Opus 5.5Anthropic	1M	128K	$4.00	$20.00	$0.20	93	Proprietary20% cheaper per token than Opus 5 ($5/$25). Cache reads are 0.05x base input ($0.20). Fast mode (research preview, Claude API only): $8 input / $40 output. Uses the newer tokenizer (about 30% more tokens than pre-Opus 4.7 models). Safety classifiers can refuse; since 24 Sep 2026 some pre-output refusal categories are billed. Minimum cacheable prompt: 512 tokens.
Claude Sonnet 5.5Anthropic	1M	128K	$2.00	$10.00	$0.20	139	ProprietarySame per-token price as Sonnet 5. Cache reads use the standard 0.1x multiplier ($0.20). Uses the newer tokenizer (about 30% more tokens than pre-Opus 4.7 models). Safety classifiers can refuse; since 24 Sep 2026 some pre-output refusal categories are billed. Minimum cacheable prompt: 512 tokens.
Claude Fable 5Anthropic	1M	128K	$10.00	$50.00	$1.00	–	ProprietaryLegacy; Fable 5.1 has the same input/output price but cache reads at $0.25 instead of $1. Uses the newer tokenizer (about 30% more tokens than pre-Opus 4.7 models). Safety classifiers can refuse; since 24 Sep 2026 some pre-output refusal categories are billed. Minimum cacheable prompt: 512 tokens.
Claude Mythos 5Anthropic	1M	128K	$10.00	$50.00	$1.00	–	ProprietaryInvite onlyInvite only through Project Glasswing. Same prices as Claude Fable 5. Uses the newer tokenizer. Minimum cacheable prompt: 512 tokens.
Claude Opus 5Anthropic	1M	128K	$5.00	$25.00	$0.50	55	ProprietaryLegacy; Opus 5.5 is cheaper ($4/$20). Fast mode (research preview, Claude API only): $10 input / $50 output. Uses the newer tokenizer. Safety classifiers can refuse; since 24 Sep 2026 some pre-output refusal categories are billed. Minimum cacheable prompt: 512 tokens.
Claude Sonnet 5Anthropic	1M	128K	$2.00	$10.00	$0.20	85	Proprietary$2/$10 launched as introductory pricing through 31 Aug 2026 and is now the standard price; the planned rise to $3/$15 on 1 Sep 2026 did not happen. Uses the newer tokenizer (launch post: roughly 1.0–1.35x more tokens depending on content). Minimum cacheable prompt: 1,024 tokens.
Claude Opus 4.8Anthropic	1M	128K	$5.00	$25.00	$0.50	61	ProprietaryLegacy. Fast mode (research preview, Claude API only): $10 input / $50 output. Uses the newer tokenizer. Minimum cacheable prompt: 1,024 tokens.
Claude Opus 4.7Anthropic	1M	128K	$5.00	$25.00	$0.50	48	ProprietaryLegacy. First model on the newer tokenizer (launch post: roughly 1.0–1.35x more tokens for the same input). Fast mode is not available. Minimum cacheable prompt: 2,048 tokens.
Claude Opus 4.6Anthropic	1M	128K	$5.00	$25.00	$0.50	38	ProprietaryLegacy. Uses the previous tokenizer. Fast mode requests run at standard speed and standard rates. Minimum cacheable prompt: 4,096 tokens.
Claude Sonnet 4.6Anthropic	1M	128K	$3.00	$15.00	$0.30	44	ProprietaryLegacy. Uses the previous tokenizer, so it counts fewer tokens for the same text than Sonnet 5 and 5.5. Minimum cacheable prompt: 1,024 tokens.
Qwen3.8-MaxAlibaba	1M	131K	$2.00	$6.00	$0.25	39	Open weightInternational (Singapore) list price. Implicit cache $0.25; explicit cache creation $2.50, explicit cache read $0.17 (Qwen Cloud). Global deployment scope is cheaper at $1.65 / $4.951. A faster qwen3.8-max-prime mode costs about 2x. Snapshot qwen3.8-max-0902 added 2 Sep 2026.
Qwen3.7-MaxAlibaba	1M	131K	$2.50	$7.50	$0.50	203	ProprietaryInternational (Singapore) list price. Implicit cache $0.50; explicit cache read $0.25. Global deployment scope $1.65 / $4.951. Alibaba now lists Qwen3.7 under Legacy models; Qwen3.8-Max is cheaper ($2 / $6).
Qwen3.8-27BAlibaba	1M	131K	$0.50	$3.00	$0.10	46	Open weightInternational (Singapore) price on Alibaba Model Studio / Qwen Cloud. Explicit cache read $0.05.
Qwen3.8-FlashAlibaba	1M	131K	$0.15	$0.47	$0.016	–	Open weightInternational (Singapore) price. Explicit cache read $0.016.
MiniMax-M3MiniMax	1M	524K	$0.30	$1.20	$0.06	88	Open weightPrices after a 'permanent 50% off' (list $0.60 / $2.40). Priority tier costs 1.5x standard.
Llama 4 MaverickMeta	1M	–	$0.24	$0.97	–	80	Open weightMeta does not sell Llama through its own API. Prices are from Amazon Bedrock (us-east-1, on-demand, model ID meta.llama4-maverick-17b-instruct-v1:0); Bedrock batch is $0.12 in / $0.485 out. Bedrock caps output at 8K tokens. Groq retired Llama 4 Maverick on 9 Mar 2026 and Together no longer lists it on serverless.
Grok 4.20xAI	1M	–	$1.25	$2.50	$0.20	115	ProprietarySame price for grok-4.20-0309-non-reasoning and grok-4.20-multi-agent-0309. Alias grok-4.20 points to the reasoning model. Priority processing costs 2x.
Grok 4.3xAI	1M	–	$1.25	$2.50	$0.20	132	ProprietaryGrok 3, Grok 4 (0709), Grok 4 Fast and Grok 4.1 Fast slugs now redirect here and are billed at Grok 4.3 rates. Priority processing costs 2x.
Amazon Nova 2 LiteAmazon	1M	66K	$0.30	$2.50	$0.075	–	ProprietaryAmazon Bedrock, global cross-region inference, Standard tier, us-east-1 ($0.33/$2.75 for geo cross-region and in-region). Batch and Flex tiers are $0.15/$1.25; Priority is $0.525/$4.375. Cached input derived from Bedrock's note that cache reads cost 75% less than on-demand input. Bedrock lists EOL no sooner than 2 Dec 2026.
Amazon Nova 2 Pro (Preview)Amazon	1M	–	$1.25	$10.00	$0.3125	–	ProprietaryInvite onlyAmazon Bedrock, global cross-region inference, Standard tier, us-east-1 ($1.375/$11 geo cross-region). Batch and Flex are $0.625/$5; Priority is $2.1875/$17.50. Cached input derived from the 75%-off cache-read note. Still in preview: early access for Amazon Nova Forge customers via the AWS account team.
Grok 4.7xAI	500K	–	$2.00	$6.00	$0.50	79	ProprietaryBatch API is not supported for grok-4.7. Priority processing costs 2x; the US regional endpoint costs 1.1x. A faster variant (Grok 4.7 Fast, 2x price) is only in Cursor and Grok Build, not the public API.
GPT-5.4 miniOpenAI	400K	128K	$0.75	$4.50	$0.075	213	ProprietaryThere is no separate cache-write charge. Batch and Flex: $0.375 / $2.25. Fast mode: $1.50 / $9. Snapshot gpt-5.4-mini-2026-03-17. Its sibling gpt-5.4-nano is deprecated (shutdown 1 Apr 2027, replacement gpt-6-luna).
GPT-5OpenAI	400K	128K	$1.25	$10.00	$0.125	75	ProprietaryRetires 11 Dec 2026There is no separate cache-write charge. Batch and Flex: $0.625 / $5. Fast mode: $2.50 / $20. Snapshot gpt-5-2025-08-07 shuts down 11 Dec 2026; the recommended replacement is gpt-5.6-sol.
Gemma 4 31BGoogle	262K	–	–	–	–	35	Open weightOpen weights. On the Gemini API Gemma 4 has only a free tier; the paid tier is listed as 'Not available', so there is no per-token price.
Gemma 4 26B A4BGoogle	262K	–	–	–	–	–	Open weightOpen weights. On the Gemini API Gemma 4 has only a free tier; the paid tier is listed as 'Not available'.
Qwen3-MaxAlibaba	262K	66K	$1.20	$6.00	$0.24	63	ProprietaryRetires 10 Oct 2026International (Singapore) price for prompts up to 32K. Alias points to qwen3-max-2026-01-23. Explicit cache read $0.12. Qwen Cloud marks it for retirement on 10 Oct 2026 (UTC+8).
Kimi K2.6Moonshot	262K	–	$0.95	$4.00	$0.16	68	Open weightPrices exclude taxes. Cache hits are automatic.
Tencent Hy3Tencent	262K	128K	$0.132	$0.528	$0.033	87	Open weightTencent Cloud TokenHub, Singapore region. OpenRouter now lists the same $0.132 / $0.528. An archived OpenRouter page from 8 Jul 2026 showed $0.20 / $0.80 plus a $0 endpoint; no free endpoint is listed now.
Mistral Medium 3.5Mistral	256K	–	$1.50	$7.50	$0.15	169	Open weightMistral API standard tier. Batch processing is 50% off per Mistral's pricing FAQ.
Mistral Large 3Mistral	256K	–	$0.50	$1.50	$0.05	79	Open weightMistral API standard tier. Batch processing is 50% off per Mistral's pricing FAQ. Much cheaper than the old Mistral Large 2 ($2/$6), which retired on 31 May 2026.
Mistral Small 4Mistral	256K	–	$0.15	$0.60	$0.015	183	Open weightMistral API standard tier. Batch processing is 50% off per Mistral's pricing FAQ.
Claude Haiku 4.5Anthropic	200K	64K	$1.00	$5.00	$0.10	91	ProprietaryCheapest current Claude model. Uses the previous tokenizer. Minimum cacheable prompt: 4,096 tokens. Retirement commitment is only 'not sooner than 15 Oct 2026'; no deprecation notice yet.
Claude Opus 4.5Anthropic	200K	64K	$5.00	$25.00	$0.50	46	ProprietaryLegacy. Uses the previous tokenizer. Minimum cacheable prompt: 4,096 tokens. Retirement commitment: not sooner than 24 Nov 2026.
Claude Sonnet 4.5Anthropic	200K	64K	$3.00	$15.00	$0.30	–	ProprietaryRetires 30 Nov 2026Deprecated 30 Sep 2026; retires on the Claude API on 30 Nov 2026. Recommended replacement: Claude Sonnet 5.5 ($2/$10). Uses the previous tokenizer. Minimum cacheable prompt: 1,024 tokens.
o3OpenAI	200K	100K	$2.00	$8.00	$0.50	102	ProprietaryRetires 11 Dec 2026Batch: $1 / $4. Flex: $1 / $4 (cached $0.25). Fast mode: $3.50 / $14. Snapshot o3-2025-04-16 shuts down 11 Dec 2026; the recommended replacement is gpt-5.6-sol. o3-pro ($20 / $80) shuts down on the same date.
o4-miniOpenAI	200K	100K	$1.10	$4.40	$0.275	142	ProprietaryRetires 23 Oct 2026Shuts down 23 Oct 2026 (both o4-mini and o4-mini-2025-04-16); the recommended replacement is gpt-5.6-terra. Batch: $0.55 / $2.20. Flex: $0.55 / $2.20 (cached $0.138). Fast mode: $2 / $8.
GLM-4.6Z.ai	200K	131K	$0.60	$2.20	$0.11	35	Open weightCached-input storage is free for a limited time. Same price as GLM-4.7.
gpt-oss-120bOpenAI	131K	131K	–	–	–	164	Open weightOpen weights on Hugging Face (openai/gpt-oss-120b). Not on OpenAI's API pricing page; the price depends on the host. Runs on a single 80 GB GPU (MXFP4).
gpt-oss-20bOpenAI	131K	131K	–	–	–	180	Open weightOpen weights on Hugging Face (openai/gpt-oss-20b). Not on OpenAI's API pricing page; the price depends on the host. Runs on devices with 16 GB of memory.
Codestral 25.08Mistral	128K	–	$0.30	$0.90	$0.03	–	ProprietaryMistral API standard tier. Batch processing is 50% off per Mistral's pricing FAQ.
Command A+Cohere	128K	64K	–	–	–	211	Open weightCohere publishes no per-token rate. Its pricing page lists Command A+ as "Free" ($0 API key, $0 model download); production rate limits are "contact sales", and dedicated hosting is via Model Vault. Third-party sites quote other prices that Cohere does not show.

Benchmark glossary

What each benchmark measures, and why it matters for picking a model.

Terminal-Bench 4.0

Sixty-six fresh tasks done in a sandboxed terminal, from software and ML to security and science. The median task is about four hours of expert work.

tbench.ai

GPQA Diamond

Expert-written biology, physics and chemistry questions built so that a web search does not give the answer away. PhD experts score around 65%.

arxiv.org/abs/2311.12022

DeepSWE v1.1

Tasks written from scratch by engineers in active open-source repositories across five languages, graded by hand-written behaviour tests. Built to resist memorised answers.

deepswe.datacurve.ai

Humanity’s Last Exam (no tools)

Questions written by subject experts at the edge of human knowledge, answered from the model alone. The hardest general-knowledge test in wide use.

lastexam.ai

Humanity’s Last Exam (with tools)

The same expert questions, with the model allowed to search the web and run code. Closer to how a research agent works.

lastexam.ai

OSWorld 2.0

Computer-use agents work through long real workflows across web apps and desktop software. We show the partial-credit score, which counts checkpoints reached; labs set the test up in slightly different ways.

os-world.github.io

SWE-bench Pro

Scale AI’s harder follow-up to SWE-bench: longer tasks in larger codebases where a fix usually touches several files. We use the public set.

scale.com/leaderboard

SWE-bench Verified

Real issues from popular Python repositories, each checked by a person to be solvable. The patch counts only if the project’s own tests pass. Mostly quoted for 2025 models now.

swebench.com

Terminal-Bench 2.1

The earlier Terminal-Bench task set. Scores are not comparable with 4.0, which uses new and harder tasks.

tbench.ai

Terminal-Bench 2.0

The 2025 Terminal-Bench task set, still quoted for older models.

tbench.ai

AutomationBench

Zapier’s test of hundreds of real business workflows, from sales and support to finance and HR, across simulated SaaS apps with messy data.

zapier.com

MMMU-Pro

A harder version of MMMU: college-level questions that only make sense with the chart, diagram or photo, with distractors that punish guessing.

mmmu-benchmark.github.io

MRCR v2 (8-needle)

Eight near-identical requests are hidden in one very long conversation and the model must return the right one. A test of recall across a large context.

huggingface.co/datasets/openai/mrcr

FrontierCode 1.1

Cognition’s benchmark: the agent fixes a real open-source issue and the patch is graded by hidden tests and a maintainer rubric for correctness, regressions and style.

cognition.com/blog/frontier-code-1.1

ARC-AGI-2

Visual puzzles where the model must infer an unseen rule from a few examples. Built to reward generalisation over memory.

arcprize.org

AIME 2025

The 2025 American Invitational Mathematics Examination: short-answer problems that need several steps of exact reasoning.

maa.org

τ²-bench

An agent serves a simulated customer in airline, retail or telecom settings, using tools and following policy.

github.com/sierra-research/tau2-bench

BrowseComp

Questions whose answers are hard to find and easy to check, so the agent has to search, read and combine many pages.

openai.com/index/browsecomp

OSWorld-Verified

The model operates desktop apps through screenshots, mouse and keyboard to finish everyday tasks, checked by scripts.

os-world.github.io

CharXiv Reasoning

Questions about real charts taken from arXiv papers, where the answer needs reasoning over the figure, not just reading a label.

charxiv.github.io

LiveCodeBench

Programming problems collected after a model’s training cut-off, so the model cannot have seen the answers.

livecodebench.github.io

Picking a model for a product?

We test models on your prompts, not on benchmarks.

A free call: we look at your workload, run cost-to-quality checks across providers and tell you which model to build on. No lock-in.

Book a free call Browse free models

Questions

AI model leaderboard FAQ

Which AI model scores highest in 2026?

It depends on the work. On Terminal-Bench 4.0, the coding benchmark Anthropic, OpenAI and Google all report, the top published score is Claude Sonnet 5.5 (70.6%). On GPQA Diamond it is GPT-6 Astra (96%). Pick the benchmark closest to your work in the chart above.

What is an AI model leaderboard?

A ranking of models by public benchmark scores, price, context window and speed, so you can shortlist without reading every provider’s documentation.

How do I compare AI models side by side?

Tick up to four models in the table and press Compare, or open the Compare tab and search for them. The page link keeps your picks, so you can send it to your team.

Which AI model API is the cheapest?

By input price, GPT-6 Luna at $0.10 per million tokens. By output price, Qwen3.8-Flash at $0.47. Which is cheaper for you depends on your input to output mix; the pricing index works it out at your own volume.

Why do some cells show a dash?

We only show numbers we could trace to a source: the provider’s own pricing page or docs, the lab’s launch post, or the benchmark’s leaderboard. A dash means the number is not published, never that it is zero.

Why are some popular benchmarks missing for some models?

Labs report different benchmarks, and in 2026 Anthropic stopped reporting SWE-bench Verified and GPQA Diamond. We never mix versions of a test (Terminal-Bench 2.1 and 4.0 are different task sets), so a model only appears in a ranking for the exact test it reported.

Compare

Compare models side by side

Pricing index

What each model costs at your volume

Free models

Free AI models and APIs

Calculator

LLM cost with caching

How we verify. Prices come from each provider’s pricing page and scores from the lab’s launch post, model card or the benchmark’s own leaderboard. Each model page lists its sources and the date we checked them (2 Oct 2026). Speed is measured by Artificial Analysis.

No affiliate links. HorizonLux has no commercial relationship with any provider listed here. Rankings are computed, not sold.