Artificial intelligence / DATED RESEARCH SNAPSHOT

Which AI models are available now?

Compare selected current API and open-weight model releases with benchmark evidence, price, and serving metadata.

Snapshot accessed 2026-09-23
11 reviewed model/version rows · See provider coverage and review gaps
Release dates retain source precision; a page update is not treated as a release date.

Observations

These findings describe the full dated snapshot. Filtering records below does not change the analysis.

Evidence is incomplete. 7 of 11 models have retained numeric evaluation results, and 9 have both token prices. Missing evidence is unknown, not zero performance; coverage reflects our retained sources.
Listed input prices range from $0.10 to $10.00. Across 9 priced models, per million input tokens. GPT-6 Luna is the lower endpoint; GPT-6 Astra is an upper endpoint. Output costs are separate, and price alone does not establish quality or total task cost.
Scoring rules change the result by 36.2 points. Claude Fable 5.1 reports 77.9% partial completion and 41.7% strict completion on OSWorld 2.0. These are different scoring rules, not a change in model ability over time. Source: anthropic-fable51 in References.
3 of 11 models have no verified maximum output in this snapshot. Check the limits below before choosing a model for long responses. A context window is a shared budget, not a promise of equally long output.
Latency is missing for 11 of 11 models. The snapshot cannot tell you which is fastest. Test your own workload, concurrency and completion criteria.

PROVIDER / MODEL-CARD RESULTS · PERCENT

Reported benchmark results

25 results · 7 models

Every retained percentage is shown below, grouped by model. Each row names its benchmark; tasks, scoring, tools and settings differ even when benchmark names match. Bar lengths show reported percentages, not a universal quality ranking. Models without retained scores are not assigned zero.

On smaller screens, scroll the chart horizontally to read every label and value.

View all reported evaluation data

Accessible data table, reported benchmark percentages
Model	Provider	Benchmark/version	Score	Evaluation setting	Source IDs
GLM-5.3-Flash	Z.ai	Terminal-Bench 2.1	84.3%	Claude Code 2.1.207; temperature 1; top_p 1; max generation 65,536; 6-hour limit	glm53-card
GLM-5.3-Flash	Z.ai	DeepSWE v1.1	63.4%	mini-swe-agent; temperature .95; top_p 1; 400K context (not output); 6-hour limit	glm53-card
GLM-5.3-Flash	Z.ai	Humanity's Last Exam (with tools)	55.3%	Full set; temperature 1; top_p .95; max generation 163,840; 300K evaluation context; GPT-5.6 Luna medium judge	glm53-card
Kimi K3	Moonshot AI	GPQA Diamond	93.5%	Max reasoning; temperature 1; top_p .95; single-step evaluation	kimi-card
Kimi K3	Moonshot AI	Terminal-Bench 2.1	88.3%	Max reasoning; temperature 1; top_p 1; Kimi Code harness	kimi-card
Kimi K3	Moonshot AI	DeepSWE v1.1	67.5%	Max reasoning; temperature 1; top_p 1; Kimi Code harness (not mini-SWE-agent)	kimi-card
GPT-6 Sol	OpenAI	ExploitBench (Cap Percent)	81.7%	OpenAI system card section 11.8.1.2.1, Figure 79; maximum reasoning effort. Capability-flag credit across five attempts on 41 vulnerabilities, not simple pass@1. Historical-vulnerability contamination warning applies.	openai-sol-luna-system-card
GPT-6 Sol	OpenAI	ExploitBench - Internal Port (June-August 2026)	5.5%	OpenAI system card section 11.8.1.2.2, Figure 80; maximum reported success rate on recent vulnerabilities. Exact effort associated with this maximum not verified from figure.	openai-sol-luna-system-card
GPT-6 Sol	OpenAI	SEC-Bench Pro	66.3%	OpenAI system card section 11.8.1.2.2, Figure 81; vulnerability discovery in JavaScript engines. Exact effort and output-token budget for this point not verified.	openai-sol-luna-system-card
GPT-6 Sol	OpenAI	ExploitGym	22.1%	OpenAI system card section 11.8.1.2.2, Figure 82; challenges with successful exploit development. Exact effort and token budget not verified; not a general coding accuracy measure.	openai-sol-luna-system-card
GPT-6 Luna	OpenAI	ExploitBench (Cap Percent)	43.4%	OpenAI system card section 11.8.1.2.1, Figure 79; maximum reasoning effort. Capability-flag credit across five attempts on 41 vulnerabilities, not simple pass@1. Historical-vulnerability contamination warning applies.	openai-sol-luna-system-card
GPT-6 Luna	OpenAI	ExploitBench - Internal Port (June-August 2026)	0.0%	OpenAI system card section 11.8.1.2.2, Figure 80; source explicitly reports no completed challenges. This is an observed zero, not missing evidence. Exact effort not verified.	openai-sol-luna-system-card
GPT-6 Luna	OpenAI	SEC-Bench Pro	34.2%	OpenAI system card section 11.8.1.2.2, Figure 81; vulnerability discovery in JavaScript engines. Exact effort and output-token budget for this point not verified.	openai-sol-luna-system-card
GPT-6 Luna	OpenAI	ExploitGym	11.6%	OpenAI system card section 11.8.1.2.2, Figure 82; challenges with successful exploit development. Exact effort and token budget not verified; not a general coding accuracy measure.	openai-sol-luna-system-card
Claude Opus 5.5	Anthropic	Terminal-Bench 4.0	66.4%	Anthropic; Claude Code --bare; adaptive xhigh; 5 trials/task; safeguards and fallback enabled. Detailed card: 66.36%, standard error 2.6 points.	anthropic-opus55-launch, anthropic-opus55-card
Claude Opus 5.5	Anthropic	Humanity's Last Exam (no tools)	64.4%	Anthropic; max effort; no tools; 1M total-token cap; no compaction; Opus 4.6 grader.	anthropic-opus55-card
GPT-6 Astra	OpenAI	Terminal-Bench 4.0	57.9%	OpenAI evaluation; provider setting retained in launch report	openai-gpt6-launch
GPT-6 Astra	OpenAI	GPQA Diamond	96.0%	OpenAI evaluation; provider setting retained in launch report	openai-gpt6-launch
Claude Fable 5.1	Anthropic	Terminal-Bench-Science 0.1	52.6%	Production safeguards enabled; Anthropic setup	anthropic-fable51
Claude Fable 5.1	Anthropic	Terminal-Bench 4.0	55.8%	Production safeguards enabled; Anthropic setup	anthropic-fable51
Claude Fable 5.1	Anthropic	Humanity's Last Exam (no tools)	60.9%	Production safeguards enabled; no tools	anthropic-fable51
Claude Fable 5.1	Anthropic	Humanity's Last Exam (with tools)	65.0%	Production safeguards enabled; with tools	anthropic-fable51
Claude Fable 5.1	Anthropic	CursorBench 3.2.0	73.4%	Max effort; Anthropic report	anthropic-fable51
Claude Fable 5.1	Anthropic	OSWorld 2.0 (partial)	77.9%	August 2026 task release; partial completion; production safeguards enabled	anthropic-fable51
Claude Fable 5.1	Anthropic	OSWorld 2.0 (strict)	41.7%	August 2026 task release; strict completion; production safeguards enabled	anthropic-fable51

LISTED TOKEN PRICE · USD / 1M TOKENS

Input and output price

metadata, not quality

Uncached input and output rates per one million tokens. Verification dates differ by model and appear in the pricing tiers table below. Each row names its tier. Gemini uses its introductory rate; DeepSeek uses peak pricing as an explicit conservative scenario. All retained tiers and cached-input rates appear in the pricing details below.

Accessible data table, listed price in USD per one million tokens
Model	Provider	Input	Output	Selected tier
Kimi K3	Moonshot AI	$3.00	$15.00	Standard Kimi API, excluding cache writes
Gemini 3.8 Flash	Google	$0.75	$3.75	Introductory, through 2026-12-31
GPT-6 Sol	OpenAI	$2.00	$10.00	Standard, prompts up to 272K input
GPT-6 Luna	OpenAI	$0.10	$0.50	Standard, prompts up to 272K input
Claude Opus 5.5	Anthropic	$4.00	$20.00	Standard Claude API, global routing
DeepSeek V4.1-Flash	DeepSeek	$0.30	$1.20	Peak hours, conservative scenario
GPT-6 Astra	OpenAI	$10.00	$50.00	Standard, prompts up to 272K input
Claude Fable 5.1	Anthropic	$10.00	$50.00	Standard Claude API
Mistral Large 3	Mistral AI	$0.50	$1.50	Standard Mistral API

Which models fit your workload?

Compare supported inputs, response limits and hosting options before testing quality. Limits retain each source’s definition; unknown means unverified, not unsupported. Open weights can carry license restrictions and require substantial hardware.

On smaller screens, scroll the table horizontally to compare all fields.

Capabilities and documented limits for every reviewed model
Model	Supported inputs and features	Context or input limit	Maximum output	Hosting / availability	Evidence
Claude Fable 5.1	Text and image input; text output; adaptive thinking	1M-token context window	128K tokens	Proprietary API model	Source Verified 2026-09-19
Claude Opus 5.5	Text and image input; text output; always-on adaptive thinking, default medium effort; tools and structured outputs. Thinking cannot be disabled; forced tool_choice any/tool is unsupported.	1M-token context window	128K tokens; Message Batches API beta supports up to 300K with output-300k-2026-03-24	Proprietary API model	Source Verified 2026-09-23
DeepSeek V4.1-Flash	Text and image input; text output; thinking and non-thinking; JSON output; tool calls	1M-token context window	384K tokens	Open-weight status indicated by provider; license not verified in retained source	Source Verified 2026-09-19
GLM-5.3-Flash	Text and image input; text output; low/high/max thinking effort	Not published in retained source	Not published in retained source	Open-weight and model materials under MIT	Source Verified 2026-09-19
GPT-6 Astra	Text and image input; text output; function calling; structured output	1,050,000-token context window (not an input-only limit)	128,000 tokens	Proprietary API model	Source Verified 2026-09-19
GPT-6 Luna	Text and image input; text output; streaming; structured outputs; Responses tools and function calling. Reasoning: none, low, medium (default), high, xhigh, max. Chat Completions function calling requires reasoning_effort none. No native audio/video or fine-tuning.	1,050,000-token context window; maximum input 922,000 tokens	128,000 tokens	Proprietary API model	Source Verified 2026-09-23
GPT-6 Sol	Text and image input; text output; streaming; structured outputs; Responses tools and function calling. Reasoning: none, low, medium (default), high, xhigh, max. Chat Completions function calling requires reasoning_effort none. No native audio/video or fine-tuning.	1,050,000-token context window; maximum input 922,000 tokens	128,000 tokens	Proprietary API model	Source Verified 2026-09-23
Gemini 3.8 Flash	Text, image, audio, video and PDF input; text output; function calling; structured output	1,048,576-token input limit	65,536 tokens	Proprietary API model	Source Verified 2026-09-19
Kimi K3	Text, image and video input; text output; always-on thinking; tool calls; strict JSON Schema. Custom license: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE	1,048,576-token context window	max_completion_tokens up to 1,048,576; default 131,072; subject to shared context budget	Open-weight; custom license	Source Verified 2026-09-19
Mistral Large 3	Text and image input; text output; function calling; structured output	256K-token context window	Not published in retained source	Open-weight	Source Verified 2026-09-19
Qwen3.8 2.4T A95B	Text input and output. Hosted Qwen3.8-Max adds capabilities not assigned to these weights.	262,144 native / 1,010,000 extensible tokens	Not published in retained source	Open-weight; license is not OSI-identified in retained source	Source Verified 2026-09-19

Pricing tiers and cache reads

USD per million tokens. Cache-read prices apply only to eligible cached input and do not include cache creation/storage. The price chart uses uncached input. Unknown is not free. Provider-direct API rates can differ from cloud resellers or subscriptions. Pricing is a dated source check, not continuously monitored.

Model	Tier	Uncached input	Cached input	Output	Conditions	Verified	Evidence
Kimi K3	Chart tier: Standard Kimi API, excluding cache writes	$3	$0.3	$15	Flat context pricing. Separate cache writes: $3/1M for 5-minute TTL (default), $6/1M for 1-hour TTL. These are excluded from the token-only cost chart; automatic caching can affect the bill. See the pricing source and usage fields.	2026-09-19	Kimi inference pricing
Gemini 3.8 Flash	Chart tier: Introductory, through 2026-12-31	$0.75	Unknown	$3.75	Temporary introductory rate; cached-input rate not retained.	2026-09-19	Gemini API latest-model guide and pricing
Gemini 3.8 Flash	Alternative: Announced standard, from 2027-01-01	$1.5	Unknown	$7.5	Future rate, excluded from the dated chart.	2026-09-19	Gemini API latest-model guide and pricing
GPT-6 Sol	Chart tier: Standard, prompts up to 272K input	$2	$0.2	$10	Cache writes $2.50/1M. Above 272K input, full-request long-context rates apply. Batch/Flex cost 50% of applicable Standard rates; Fast costs 2x. Eligible regional processing adds 10%; EU data residency requires Standard. Tool charges are separate. Effective pricing interval not expressly dated; observed September 23.	2026-09-23	GPT-6 Sol; Pricing
GPT-6 Sol	Alternative: Standard, prompts above 272K input	$4	$0.4	$15	Applies to the entire request, not only excess tokens. Cache writes $5/1M. Other processing tiers and regional premiums remain separate.	2026-09-23	Pricing
GPT-6 Luna	Chart tier: Standard, prompts up to 272K input	$0.1	$0.01	$0.5	Cache writes $0.125/1M. Above 272K input, full-request long-context rates apply. Batch/Flex cost 50% of applicable Standard rates; Fast costs 2x. Eligible regional processing adds 10%; EU data residency requires Standard. Tool charges are separate. Effective pricing interval not expressly dated; observed September 23.	2026-09-23	GPT-6 Luna; Pricing
GPT-6 Luna	Alternative: Standard, prompts above 272K input	$0.2	$0.02	$0.75	Applies to the entire request, not only excess tokens. Cache writes $0.25/1M. Other processing tiers and regional premiums remain separate.	2026-09-23	Pricing
Claude Opus 5.5	Chart tier: Standard Claude API, global routing	$4	$0.2	$20	Full 1M context at standard token rates. Separate cache writes: $5/1M for 5m and $8/1M for 1h. Tool charges, residency premiums and subscriptions excluded; cloud partner prices can differ.	2026-09-23	Claude Opus 5.5; Pricing
Claude Opus 5.5	Alternative: Claude Message Batches API	$2	Unknown	$10	Asynchronous batch processing. Cache-read rate not retained in this tier; null is unknown, not free. Not combinable with fast mode.	2026-09-23	Pricing
Claude Opus 5.5	Alternative: Claude API fast mode, research preview	$8	Unknown	$40	First-party Claude API only, speed fast and fast-mode-2026-02-01 beta header. Cache-read rate not retained here; caching and residency modifiers apply separately. Not a separate model release.	2026-09-23	Pricing
DeepSeek V4.1-Flash	Chart tier: Peak hours, conservative scenario	$0.3	$0.006	$1.2	Peak: weekdays 01:00-04:00 and 06:00-10:00 UTC, excluding Chinese public holidays. Chart uses peak regardless of viewing time.	2026-09-19	Models and pricing
DeepSeek V4.1-Flash	Alternative: Off-peak hours	$0.15	$0.003	$0.6	All other hours, including weekends and Chinese public holidays.	2026-09-19	Models and pricing
GPT-6 Astra	Chart tier: Standard, prompts up to 272K input	$10	$1	$50	Above 272K input tokens, the full request uses 2x input/cache rates and 1.5x output rates. Cache writes cost $12.50/1M at standard context. Batch/Flex and Fast use different rates.	2026-09-19	GPT-6 Astra model documentation
Claude Fable 5.1	Chart tier: Standard Claude API	$10	$0.25	$50	Cache writes: $12.50/1M for 5 minutes, $20/1M for 1 hour. Batch pricing is separate.	2026-09-19	Claude Fable 5.1 specifications
Mistral Large 3	Chart tier: Standard Mistral API	$0.5	$0.05	$1.5	Cached rate is a cache read, not an uncached input price. Batch and priority pricing are separate.	2026-09-19	Inference pricing

About the data

Coverage and models awaiting review

11 providers tracked; 3 have no reviewed comparison yet. 0 source checks failed; 2 scans reached their limit. Catalogs checked 2026-09-23 14:27 UTC. A catalog check discovers candidates; it does not verify their specifications, prices or availability. 0 published models are due for a new fact check.

Starting discovery register, not an exhaustive provider or model inventory. Reviewed comparisons focus on text, coding and multimodal language models. Catalogs can contain old, retired or specialist models. Unreviewed counts are research candidates, not a count of missing current releases. Catalog discovery runs on each pipeline collection, not continuously. Model facts require separate source review; older than 30 days is flagged for review.

View provider coverage and review gaps

Provider source	Reviewed models	Discovery status	Unreviewed candidates
OpenAI	GPT-6 Astra, GPT-6 Luna, GPT-6 Sol	Catalog checked	18
Anthropic	Claude Fable 5.1, Claude Opus 5.5	Catalog checked	3
Google	Gemini 3.8 Flash	Catalog checked	41
Qwen	Qwen3.8 2.4T A95B	Scan limit reached	149
Z.ai	GLM-5.3-Flash	Scan limit reached	149
Mistral AI	Mistral Large 3	Catalog checked	74
DeepSeek	DeepSeek V4.1-Flash	Catalog checked	104
Moonshot AI	Kimi K3	Catalog checked	18
Meta	None reviewed yet	Catalog checked	70
xAI	None reviewed yet	Catalog checked	3
Cohere	None reviewed yet	Catalog checked	11

Official Hugging Face organizations: up to 3 pages of 50 entries sorted by last modification. HTML catalogs: matching model links on the configured page. Not a complete market census.

Download candidate IDs, source checks and review gaps

Shortlist by the work you need to do

Worked example: for a video-summary workflow, the retained Gemini 3.8 Flash and Kimi K3 records explicitly document video input. Check the provider’s file format and upload requirements, then test representative clips. Their large input/context allowances do not guarantee the same response length: Gemini lists a 65,536-token output limit, while Kimi documents a generation parameter subject to the shared context budget.

For long documents, budget for instructions, document content, conversation history and generated output. For private deployment, inspect the exact weight license and hardware needs. Absence of a feature in this table is not a confirmed lack of support. The comparison narrows what to test; benchmark percentages do not establish success on your own task.

All reported evaluations are visible together, grouped by model. Each benchmark has its own tasks, scoring rules and run settings. Neither the chart order nor the number of retained evaluations ranks model quality. Record filters below do not change these full-snapshot views.

Audit corrections, 2026-09-19: output limits are now separate from context/input limits. GPT-6 Astra has a 128,000-token output maximum; GLM evaluation context is no longer presented as a model limit. Qwen3.8-Max hosted benchmark scores were removed from the distinct open-weight record. Prices now come from validated numeric tiers, not the first dollar amount in a text field.

Coverage is selected, not exhaustive. Catalog discovery now checks a defined provider register, while only individually verified records enter this comparison; Mistral Large 3 is included as a dated open-weight large-model reference and is not presented as a claim about every newer Mistral release. “Open-weight” means weights or self-hosting evidence were retained; it does not automatically mean an OSI-approved open-source license. The performance chart shows all retained evaluations, grouped by model. Benchmark definitions, settings, and harnesses differ, so it does not claim a universal model ranking. Prices, latency, and throughput are serving metadata and are not quality scores.

4 row(s) have no numeric benchmark retained from the listed sources. Missing values stay visible as “Not published in retained source”.

Source ledger: provenance download · research/ai-models-snapshot-2026-09-23.json · methodology

Possible explanations

Tentative, evaluation design changes the result. Partial-credit scoring rewards progress that strict completion does not. That definition is consistent with the OSWorld score gap, but this snapshot cannot isolate the effects of tools, safeguards, or run settings. A matched rerun with the same model and harness would test those effects.

Hypothesis, token prices reflect different serving choices. Infrastructure costs, introductory pricing, and product positioning could contribute to the spread. The price list alone does not reveal provider costs or performance per dollar. Test a fixed task set with measured tokens, successful outcomes, latency, and the applicable billing tier before estimating value.

Explore the records

Download all records (CSV)

ProviderAvailability

11 records

Model	Provider	Version	Release	API / serving	Openness	License	Parameters	Context / input limit	Maximum output	Capabilities	Input price	Output price	Latency	Throughput	Reported benchmark evidence	Evidence
GLM-5.3-Flash	Z.ai	GLM-5.3-Flash	Not stated in retained source	Z.ai API; self-hostable via vLLM/SGLang/Transformers	Open-weight and model materials under MIT	MIT	320B total / 18B active	Not published in retained source	Not published in retained source	Text and image input; text output; low/high/max thinking effort	Not published in retained source	Not published in retained source	Not published in retained source	Not published in retained source	View 3 results for GLM-5.3-Flash Terminal-Bench 2.1: 84.3% Model-card reported Setting: Claude Code 2.1.207; temperature 1; top_p 1; max generation 65,536; 6-hour limit DeepSWE v1.1: 63.4% Model-card reported Setting: mini-swe-agent; temperature .95; top_p 1; 400K context (not output); 6-hour limit Humanity's Last Exam (with tools): 55.3% Model-card reported Setting: Full set; temperature 1; top_p .95; max generation 163,840; 300K evaluation context; GPT-5.6 Luna medium judge Record source IDs: glm53-card	Evidence
Qwen3.8 2.4T A95B	Qwen	Qwen3.8-2.4T-A95B	Not stated in retained source	Self-hostable weights; hosted Qwen3.8-Max is a distinct serving variant	Open-weight; license is not OSI-identified in retained source	Qwen3.8 Max custom license	2.4T total / 95B active	262,144 native / 1,010,000 extensible tokens	Not published in retained source	Text input and output. Hosted Qwen3.8-Max adds capabilities not assigned to these weights.	Not published in retained source	Not published in retained source	Not published in retained source	Not published in retained source	Not published in retained source	Evidence
Kimi K3	Moonshot AI	kimi-k3	Not stated in retained source	Kimi API (OpenAI/Anthropic compatible); self-hostable weights	Open-weight; custom license	Kimi K3 custom license; commercial conditions apply	2.8T total / 104B active	1,048,576-token context window	max_completion_tokens up to 1,048,576; default 131,072; subject to shared context budget	Text, image and video input; text output; always-on thinking; tool calls; strict JSON Schema. Custom license: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE	$3 / 1M uncached tokens; cache writes billed separately	$15 / 1M tokens	Not published in retained source	Not published in retained source	View 3 results for Kimi K3 GPQA Diamond: 93.5% Model-card reported Setting: Max reasoning; temperature 1; top_p .95; single-step evaluation Terminal-Bench 2.1: 88.3% Model-card reported Setting: Max reasoning; temperature 1; top_p 1; Kimi Code harness DeepSWE v1.1: 67.5% Model-card reported Setting: Max reasoning; temperature 1; top_p 1; Kimi Code harness (not mini-SWE-agent) Record source IDs: kimi-card, kimi-guide, kimi-pricing	Evidence
Gemini 3.8 Flash	Google	gemini-3.8-flash	Not stated in retained source	Gemini API	Proprietary API model	Google API terms	Not published in retained source	1,048,576-token input limit	65,536 tokens	Text, image, audio, video and PDF input; text output; function calling; structured output	$0.75 / 1M uncached tokens (Introductory, through 2026-12-31); $1.5 / 1M uncached tokens (Announced standard, from 2027-01-01)	$3.75 / 1M tokens (Introductory, through 2026-12-31); $7.5 / 1M tokens (Announced standard, from 2027-01-01)	Not published in retained source	Not published in retained source	Not published in retained source	Evidence
GPT-6 Sol	OpenAI	gpt-6-sol	2026-09-22	OpenAI API (Responses, Chat Completions, Batch); Amazon Bedrock (Runtime and Mantle; Mantle us-east-1)	Proprietary API model	Provider terms	Not published in retained source	1,050,000-token context window; maximum input 922,000 tokens	128,000 tokens	Text and image input; text output; streaming; structured outputs; Responses tools and function calling. Reasoning: none, low, medium (default), high, xhigh, max. Chat Completions function calling requires reasoning_effort none. No native audio/video or fine-tuning.	$2 / 1M uncached tokens (Standard, prompts up to 272K input)	$10 / 1M tokens (Standard, prompts up to 272K input)	Not published in retained source	Not published in retained source	View 4 results for GPT-6 Sol ExploitBench (Cap Percent): 81.7% Provider reported Setting: OpenAI system card section 11.8.1.2.1, Figure 79; maximum reasoning effort. Capability-flag credit across five attempts on 41 vulnerabilities, not simple pass@1. Historical-vulnerability contamination warning applies. ExploitBench - Internal Port (June-August 2026): 5.5% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 80; maximum reported success rate on recent vulnerabilities. Exact effort associated with this maximum not verified from figure. SEC-Bench Pro: 66.3% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 81; vulnerability discovery in JavaScript engines. Exact effort and output-token budget for this point not verified. ExploitGym: 22.1% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 82; challenges with successful exploit development. Exact effort and token budget not verified; not a general coding accuracy measure. Record source IDs: openai-sol-docs-0923, openai-api-changelog-0922, openai-pricing-0923, openai-sol-luna-system-card	Evidence
GPT-6 Luna	OpenAI	gpt-6-luna	2026-09-22	OpenAI API (Responses, Chat Completions, Batch); Amazon Bedrock (Runtime and Mantle; Mantle us-east-1)	Proprietary API model	Provider terms	Not published in retained source	1,050,000-token context window; maximum input 922,000 tokens	128,000 tokens	Text and image input; text output; streaming; structured outputs; Responses tools and function calling. Reasoning: none, low, medium (default), high, xhigh, max. Chat Completions function calling requires reasoning_effort none. No native audio/video or fine-tuning.	$0.10 / 1M uncached tokens (Standard, prompts up to 272K input)	$0.50 / 1M tokens (Standard, prompts up to 272K input)	Not published in retained source	Not published in retained source	View 4 results for GPT-6 Luna ExploitBench (Cap Percent): 43.4% Provider reported Setting: OpenAI system card section 11.8.1.2.1, Figure 79; maximum reasoning effort. Capability-flag credit across five attempts on 41 vulnerabilities, not simple pass@1. Historical-vulnerability contamination warning applies. ExploitBench - Internal Port (June-August 2026): 0% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 80; source explicitly reports no completed challenges. This is an observed zero, not missing evidence. Exact effort not verified. SEC-Bench Pro: 34.2% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 81; vulnerability discovery in JavaScript engines. Exact effort and output-token budget for this point not verified. ExploitGym: 11.6% Provider reported Setting: OpenAI system card section 11.8.1.2.2, Figure 82; challenges with successful exploit development. Exact effort and token budget not verified; not a general coding accuracy measure. Record source IDs: openai-luna-docs-0923, openai-api-changelog-0922, openai-pricing-0923, openai-sol-luna-system-card	Evidence
Claude Opus 5.5	Anthropic	claude-opus-5-5	2026-09-22	Claude API; Amazon Bedrock; Google Cloud; Microsoft Foundry; Claude Platform on AWS	Proprietary API model	Provider terms; specific contractual reuse rights not reviewed	Not published in retained source	1M-token context window	128K tokens; Message Batches API beta supports up to 300K with output-300k-2026-03-24	Text and image input; text output; always-on adaptive thinking, default medium effort; tools and structured outputs. Thinking cannot be disabled; forced tool_choice any/tool is unsupported.	$4 / 1M uncached tokens (standard Claude API)	$20 / 1M tokens (standard Claude API)	Not published in retained source	Not published in retained source	View 2 results for Claude Opus 5.5 Terminal-Bench 4.0: 66.4% Provider reported Setting: Anthropic; Claude Code --bare; adaptive xhigh; 5 trials/task; safeguards and fallback enabled. Detailed card: 66.36%, standard error 2.6 points. Humanity's Last Exam (no tools): 64.4% Provider reported Setting: Anthropic; max effort; no tools; 1M total-token cap; no compaction; Opus 4.6 grader. Record source IDs: anthropic-opus55-spec, anthropic-opus55-launch, anthropic-opus55-card, anthropic-pricing, anthropic-platform-release-notes	Evidence
DeepSeek V4.1-Flash	DeepSeek	deepseek-flash	2026-09-10	DeepSeek API (`deepseek-flash`); provider links weights	Open-weight status indicated by provider; license not verified in retained source	Not published in retained source	552B total / 8B input-active / 16B output-active	1M-token context window	384K tokens	Text and image input; text output; thinking and non-thinking; JSON output; tool calls	$0.3 / 1M uncached tokens (Peak hours, conservative scenario); $0.15 / 1M uncached tokens (Off-peak hours)	$1.2 / 1M tokens (Peak hours, conservative scenario); $0.6 / 1M tokens (Off-peak hours)	Not published in retained source	Not published in retained source	Not published in retained source	Evidence
GPT-6 Astra	OpenAI	gpt-6-astra	2026-09-03	OpenAI API; Azure; Amazon Bedrock	Proprietary API model	Provider terms	Not published in retained source	1,050,000-token context window (not an input-only limit)	128,000 tokens	Text and image input; text output; function calling; structured output	$10 / 1M uncached tokens (Standard, prompts up to 272K input)	$50 / 1M tokens (Standard, prompts up to 272K input)	Not published in retained source	Not published in retained source	View 2 results for GPT-6 Astra Terminal-Bench 4.0: 57.9% Provider reported Setting: OpenAI evaluation; provider setting retained in launch report GPQA Diamond: 96.0% Provider reported Setting: OpenAI evaluation; provider setting retained in launch report Record source IDs: openai-api, openai-gpt6-launch	Evidence
Claude Fable 5.1	Anthropic	claude-fable-5-1	2026-09-01	Claude API; Amazon Web Services; Google Cloud; Microsoft Azure	Proprietary API model	Provider terms	Not published in retained source	1M-token context window	128K tokens	Text and image input; text output; adaptive thinking	$10 / 1M uncached tokens (Standard Claude API)	$50 / 1M tokens (Standard Claude API)	Not published in retained source	Not published in retained source	View 7 results for Claude Fable 5.1 Terminal-Bench-Science 0.1: 52.6% Provider reported Setting: Production safeguards enabled; Anthropic setup Terminal-Bench 4.0: 55.8% Provider reported Setting: Production safeguards enabled; Anthropic setup Humanity's Last Exam (no tools): 60.9% Provider reported Setting: Production safeguards enabled; no tools Humanity's Last Exam (with tools): 65.0% Provider reported Setting: Production safeguards enabled; with tools CursorBench 3.2.0: 73.4% Provider reported Setting: Max effort; Anthropic report OSWorld 2.0 (partial): 77.9% Provider reported Setting: August 2026 task release; partial completion; production safeguards enabled OSWorld 2.0 (strict): 41.7% Provider reported Setting: August 2026 task release; strict completion; production safeguards enabled Record source IDs: anthropic-fable51, claude-fable-spec	Evidence
Mistral Large 3	Mistral AI	mistral-large-2512	2025-12-02	Mistral API; Amazon Bedrock; Microsoft Azure	Open-weight	Apache 2.0	675B total / 41B active	256K-token context window	Not published in retained source	Text and image input; text output; function calling; structured output	$0.5 / 1M uncached tokens (Standard Mistral API)	$1.5 / 1M tokens (Standard Mistral API)	Not published in retained source	Not published in retained source	Not published in retained source	Evidence

Rows retain the normalized source fields. Download JSON.

References

These are the substantive provider/model-card sources used for this dated page. The record table maps benchmark settings and source IDs to the retained evidence.

GPT-6 Astra model documentation, OpenAI; data. Published/update: 2026-09-03; accessed 2026-09-19.
Introducing GPT-6 Astra, OpenAI; data. Published/update: 2026-09-03; accessed 2026-09-19.
Introducing Claude Fable 5.1 and Claude Mythos 5.1, Anthropic; data. Published/update: 2026-09-01; accessed 2026-09-19.
Gemini 3.8 Flash model documentation, Google; data. Published/update: 2026-09-17; accessed 2026-09-19.
Gemini API latest-model guide and pricing, Google; data. Published/update: 2026-09-17; accessed 2026-09-19.
Qwen3.8 2.4T A95B model card, Qwen / Hugging Face; data. Published/update: 2026-08-12; accessed 2026-09-19.
GLM-5.3-Flash model card, Z.ai / Hugging Face; data. Published/update: 2026-09-19; accessed 2026-09-19.
Mistral 3, Mistral AI; methodology/context. Published/update: 2025-12-02; accessed 2026-09-19.
Inference pricing, Mistral AI; data. Published/update: Not stated; accessed 2026-09-19.
DeepSeek V4.1-Flash, DeepSeek; data. Published/update: 2026-09-10; accessed 2026-09-19.
Claude Fable 5.1 specifications, Anthropic; data. Published/update: Not stated; accessed 2026-09-19.
Models and pricing, DeepSeek; data. Published/update: Not stated; accessed 2026-09-19.
Mistral Large 3 specifications, Mistral AI; data. Published/update: Not stated; accessed 2026-09-19.
Kimi K3 model card, Moonshot AI; data. Published/update: Not stated; accessed 2026-09-19.
Kimi K3 API guide, Moonshot AI; data. Published/update: Not stated; accessed 2026-09-19.
Kimi inference pricing, Moonshot AI; data. Published/update: Not stated; accessed 2026-09-19.
Kimi K3 license, Moonshot AI; methodology/context. Published/update: Not stated; accessed 2026-09-19.
Changelog: September 22, 2026, OpenAI; verification. Published/update: 2026-09-22; accessed 2026-09-23.
GPT-6 Sol, OpenAI; data. Published/update: Not stated; accessed 2026-09-23.
GPT-6 Luna, OpenAI; data. Published/update: Not stated; accessed 2026-09-23.
Pricing, OpenAI; verification. Published/update: Not stated; accessed 2026-09-23.
GPT-6 Astra System Card: GPT-6 Sol, GPT-6 Luna appendix, OpenAI; data. Published/update: 2026-09-22; accessed 2026-09-23.
Claude Opus 5.5, Anthropic; data. Published/update: 2026-09-22; accessed 2026-09-23.
Claude Opus 5.5, Anthropic; data. Published/update: Not stated; accessed 2026-09-23.
System Card: Claude Opus 5.5, Anthropic; methodology/context. Published/update: 2026-09-22; accessed 2026-09-23.
Pricing, Anthropic; data. Published/update: Not stated; accessed 2026-09-23.
Claude Platform release notes, Anthropic; verification. Published/update: 2026-09-22; accessed 2026-09-23.

Visualization package credits

Chart.js documentation, Chart.js; methodology/context. Published/update: Not stated; accessed 2026-09-19.
Chart.js npm package, Chart.js; methodology/context. Published/update: Not stated; accessed 2026-09-19.