Coding and engineering
Kimi K2.7 Code
$0.95 input / $4.00 output per 1M tokens
Best coding specialist to test now.
Benchmark board
Benchmarks
A benchmark win is not a buying decision. Use this board to shortlist models your team can access, govern, afford, and replace if conditions change.
Best picks
What to use now
Writing and productivity
GPT-5.5
$5.00 input / $30.00 output per 1M tokens
Best premium default for everyday professional work.
Research and reasoning
Gemini 3.1 Pro
$2.00-$4.00 input / $12.00-$18.00 output per 1M tokens
Best candidate for heavy analysis and large-context reasoning.
Image, media and documents
Split by task
Media pricing varies by model and output type.
No single public winner. Pick the media model for the job.
Cost-sensitive scale
Gemini 3.1 Flash-Lite
$0.25 input / $1.50 output per 1M tokens
Best public low-cost default to test for simple high-volume work.
Enterprise control and risk
Claude Opus 4.8
$5.00 input / $25.00 output per 1M tokens
Best Anthropic fallback while Fable is unavailable.
Agentic workflows
Gemini 3.5 Flash
$1.50 input / $9.00 output per 1M tokens
Best fast-agent candidate to test now.
Claude lane
Opus, Sonnet and Fable
Everyday Claude default
Claude Sonnet 4.6
Use for regular Claude work: summaries, drafts, coding support, document review, and supervised agents.
$3.00 input / $15.00 output per 1M tokensAvailable Anthropic defaultHard Claude work
Claude Opus 4.8
Use when the work needs the strongest available Claude option and Fable access is unavailable.
$5.00 input / $25.00 output per 1M tokensAvailable Anthropic fallbackWatchlist only
Claude Fable 5
Do not route production defaults here while access is suspended. Keep the fallback plan visible.
$10.00 input / $50.00 output per 1M tokensSuspended for all customers Jun 12Plain-English leaderboard
Best model by job
1
Coding and engineering
Use
Kimi K2.7 Code
Use it for supervised repo work, test repair, scoped refactors, and coding-agent pilots.
Cost per 1M tokens
$0.95 input / $4.00 output per 1M tokensCheap enough to pilot, but review time still drives the real cost.Watch out
Requires harness changes: keep reasoning_content during tool calls and use supported tool_choice values.2
Writing and productivity
Use
GPT-5.5
Use it when teams need one reliable model for drafting, summaries, planning, and broad knowledge work.
Cost per 1M tokens
$5.00 input / $30.00 output per 1M tokensWorth testing only where quality or review-load reduction offsets premium pricing.Watch out
Do not treat a default model as a routing policy. Track cost and fallback options.3
Research and reasoning
Use
Gemini 3.1 Pro
Use it for strategy packets, policy review, market analysis, and long documents that need careful synthesis.
Cost per 1M tokens
$2.00-$4.00 input / $12.00-$18.00 output per 1M tokensContext length changes cost; validate on real documents before routing broadly.Watch out
Best tested on internal analysis packets, not only abstract reasoning charts.4
Image, media and documents
Use
Split by task
Use Gemini 3.5 Flash for multimodal understanding; use dedicated image or video models for generation work.
Cost per 1M tokens
Media pricing varies by model and output type.Media pricing varies by model and output type. Do not compare text-token prices with image or video generation prices.Watch out
The current source set does not support naming one model as best across images, video, and documents.5
Cost-sensitive scale
Use
Gemini 3.1 Flash-Lite
Use it for simple summaries, extraction, translation, and high-volume agent steps where unit cost matters.
Cost per 1M tokens
$0.25 input / $1.50 output per 1M tokensBatch and flex modes can be lower, but quality must hold up on your own simple tasks.Watch out
Not the reasoning or coding winner; route only simple work here until internal tests prove quality.6
Enterprise control and risk
Use
Claude Opus 4.8
Use it when teams need a stable Claude option for sensitive work, deep analysis, and fallback planning.
Cost per 1M tokens
$5.00 input / $25.00 output per 1M tokensMore expensive than Sonnet or Kimi, but cheaper than building around unavailable Fable access.Watch out
Use until Fable access and data-retention terms are restored and tested.7
Agentic workflows
Use
Gemini 3.5 Flash
Use it for supervised multi-step workflows where speed, tool use, grounding, and iteration time matter.
Cost per 1M tokens
$1.50 input / $9.00 output per 1M tokensSearch and Maps grounding can add separate charges, so measure full workflow cost.Watch out
Speed is not enough for high-impact decisions; keep review states visible.Evidence usedRaw scores and source links
| Use case | Model | Metric | Result | Source |
|---|---|---|---|---|
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | GPT-5.5 | Availability | Available; verify enterprise terms | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Gemini 3.1 Pro | Availability | Available; verify Google terms | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Claude Opus 4.8 | Availability | Available Anthropic fallback | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Claude Sonnet 4.6 | Availability | Available Anthropic default | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Claude Fable 5 | Availability | Suspended for all customers Jun 12 | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Gemini 3.5 Flash | Availability | Available; verify current rollout | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Gemini 3.1 Flash-Lite | Availability | Available; low-cost Gemini API | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | DeepSeek V4 | Availability | Available through supported providers | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Kimi K2.7 Code | Availability | Released; coding-specialist API | Anthropic Fable/Mythos access statement |
Enterprise control and riskAccess status and governance risk are now part of benchmark quality because leaders need reliable defaults | Kimi K2.6 | Availability | Released; broader Kimi model | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | GPT-5.5 | Model class | General frontier | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Gemini 3.1 Pro | Model class | Reasoning and long-context | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Claude Opus 4.8 | Model class | Premium Claude fallback | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Claude Sonnet 4.6 | Model class | Balanced Claude | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Claude Fable 5 | Model class | Mythos-class, unavailable | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Gemini 3.5 Flash | Model class | Fast agentic | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Gemini 3.1 Flash-Lite | Model class | Low-cost fast model | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | DeepSeek V4 | Model class | Budget frontier alternative | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Kimi K2.7 Code | Model class | Coding specialist | Anthropic Fable/Mythos access statement |
Writing and productivityClassify the model before assigning work | Kimi K2.6 | Model class | Open-source multimodal | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | GPT-5.5 | Coding evidence | Internal harness required | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Gemini 3.1 Pro | Coding evidence | Internal harness required | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Claude Opus 4.8 | Coding evidence | Use as Anthropic default | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Claude Sonnet 4.6 | Coding evidence | Best speed/cost Claude lane | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Claude Fable 5 | Coding evidence | Strong launch claims; unavailable | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Gemini 3.5 Flash | Coding evidence | Fast agent/coding candidate | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Gemini 3.1 Flash-Lite | Coding evidence | Not a coding specialist | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | DeepSeek V4 | Coding evidence | Internal harness required | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Kimi K2.7 Code | Coding evidence | Kimi's strongest coding model | Anthropic Fable/Mythos access statement |
Coding and engineeringCoding claims should be tested on real repositories before defaults change | Kimi K2.6 | Coding evidence | SWE-bench Pro 58.6; Verified 80.2 | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | GPT-5.5 | Agentic work | Broad candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Gemini 3.1 Pro | Agentic work | Reasoning-heavy candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Claude Opus 4.8 | Agentic work | Claude workflow candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Claude Sonnet 4.6 | Agentic work | Everyday Claude agent candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Claude Fable 5 | Agentic work | Long-horizon claims; unavailable | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Gemini 3.5 Flash | Agentic work | Fast multi-step candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Gemini 3.1 Flash-Lite | Agentic work | High-volume simple agents | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | DeepSeek V4 | Agentic work | Cost-sensitive candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Kimi K2.7 Code | Agentic work | Long-context coding agent candidate | Anthropic Fable/Mythos access statement |
Agentic workflowsTool use, long-horizon work, and workflow automation need separate tests | Kimi K2.6 | Agentic work | Toolathlon 50.0; DeepSearchQA f1 92.5 | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | GPT-5.5 | Research and reasoning | Premium general reasoning | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Gemini 3.1 Pro | Research and reasoning | 94.3% | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Claude Opus 4.8 | Research and reasoning | Premium Claude reasoning | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Claude Sonnet 4.6 | Research and reasoning | Balanced Claude reasoning | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Claude Fable 5 | Research and reasoning | Strong launch claims; unavailable | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Gemini 3.5 Flash | Research and reasoning | Fast candidate | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | Gemini 3.1 Flash-Lite | Research and reasoning | Simple reasoning only | Anthropic Fable/Mythos access statement |
Research and reasoningUse public reasoning scores only as a shortlist filter | DeepSeek V4 | Research and reasoning | Budget candidate | Anthropic Fable/Mythos access statement |
Further reading
How to read the board
Scores are directional. Use them to frame workflow tests, governance checks, and buying questions.
Opus 4.8 leads model benchmark leaderboard on performance and cost efficiency
Leaderboard reshuffle: Anthropic Claude Opus 4.8 Max claims the top aggregate performance slot across 18 model configurations, but the per-task cost gap versus Cursor Composer 2.5 is 20x. Executives should not treat rank alone as a procurement signal - the spread between first and third on raw performance is 1.6 points, while the cost spread is $10.47 per task. Source URLs were not provided in the extracted structure; metric scores are drawn from stated claims in the source document. All figures should be verified against Anthropic's official release notes before operational commitment.
What AI benchmarks tell executives, and what they do not
Benchmark gains can be useful signals, but leaders should treat them as indicators of direction rather than proof of business value.
How to read benchmark comparisons before buying an AI product
A sensible buying lens for benchmark charts: look past the score and ask how the evaluation maps to your company workflows, risk tolerance, and review process.