Anthropic's flagship Opus model for long-horizon agentic coding and knowledge work, with always-on adaptive reasoning.
Modality:
Text
Image
PDF
model ID
anthropic/claude-opus-5.5
Output Speed *
97.46
tok/s
Intelligence Index *
57.6
/ 100
Context Window *
1000000
tokens
Input price
24
Anytoken
Output price
120
Anytoken
Claude Opus 5.5: Flagship-Grade Agentic Coding at a Lower Run Cost
Claude Opus 5.5 is Anthropic's flagship Opus model and the first release in the Claude 5.5 family. It targets demanding reasoning, agentic coding, and knowledge work, succeeding Claude Opus 5. Anthropic positions it as the default starting point for most workloads, reporting performance near Claude Fable 5.1 on most tasks while costing roughly 40% less to run than Opus 5 on typical workloads. Adaptive thinking is always on, steered by an effort parameter. Long-running coding agents and multi-step enterprise automation that reuse cached context benefit most.
Start building with Claude Opus 5.5 via the AnyAPI.ai API
Performance
Where Claude Opus 5.5 Earns Its Flagship Position
Opus 5.5 is built for agentic coding and long-horizon knowledge work. On Terminal-Bench 4.0, which measures multi-step professional tasks in a command line, Anthropic reports Opus 5.5 at default effort beats Opus 5 at max effort for roughly a fifth of the cost. Independent testing places it at the top of the Artificial Analysis Intelligence Index at max effort. For production, this means stronger results per token on long-running agents, with lower total cost on workloads that repeatedly reuse cached context rather than regenerate it.
Benchmarks
Independent Benchmark Signals for Claude Opus 5.5
Artificial Analysis scores Claude Opus 5.5 at 58 on its Intelligence Index at max effort, placing it at the top of the models they tested. At medium effort it scores 51, still well above the comparable-model median of 26. On agentic coding, independent Terminal-Bench 4.0 measurements rise with effort, reaching roughly 59.6% at xhigh. Measured output speed sits near 79 tokens per second at medium effort, which Artificial Analysis notes is below the reasoning-model average. Scores and speed vary meaningfully by effort level, so benchmark figures must be matched to the effort you actually deploy.
Output Speed
*
97.46
tok/s
Intelligence Index
*
57.6
/ 100
MMLU *
Broad world knowledge and problem-solving
0
%
GPQA *
PhD-level scientific reasoning across physics, biology, chemistry.
0
%
HLE *
Adherence to multi-step structured instructions.
61
%
LiveCodeBench *
Tool-calling reliability in long agentic loops.
0
%
Technical Specifications
What the model supports
Claude Opus 5.5 accepts text and image input and returns text, with a 1M-token context window and 128K maximum synchronous output (up to 300k via a Message Batches beta header). The most important production detail is reasoning: adaptive thinking is always on and cannot be disabled via the API—requests sending thinking disabled return a 400 error. Because thinking tokens count toward max_tokens and are billed as output, latency-sensitive or high-volume workloads must tune the effort parameter (low, medium, high, xhigh, max) and size max_tokens accordingly.
Verified Specifications —
Claude Opus 5.5
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1000000
tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
September 22, 2026
Pricing (standard)
24
AnyTokens in
/
120
AnyTokens out
Comparison
Claude Opus 5.5 vs Claude Opus 5: What Actually Changes?
Opus 5.5 is the direct successor to Opus 5 and the realistic upgrade decision for existing Opus users. Both share a 1M-token context window, 128K maximum output, and adaptive thinking steered by effort. The practical differences are cost, speed, and reasoning defaults: Anthropic reports Opus 5.5 costs about 40% less to run on typical workloads, with lower per-token pricing and a large cache-read discount, and generates output more than 30% faster. The default effort drops from high to medium, and thinking can no longer be disabled at all.
Dimension
Claude Opus 5.5
Claude Opus 5 (Max)
Context window *
1000000
tokens
1000000
tokens
Output speed *
97.46
tok/s
0.00
tok/s
Intelligence Index *
57.6
50.8
Input pricing
24
AnyToken
30
AnyToken
Output pricing
120
AnyToken
150
AnyToken
Knowledge cutoff *
September 22, 2026
July 2026
Choose Claude Opus 5.5 when you want higher quality per token on agentic coding and knowledge work, lower run cost on cached long-running sessions, and faster output—the default for most new workloads. Choose Claude Opus 5 only if your integration depends on disabling thinking for certain requests, since Opus 5 still accepts thinking disabled at high effort and below, whereas Opus 5.5 rejects it. Most teams should migrate, but re-run an effort sweep rather than carrying over Opus 5 settings.
Limitations & Trade-offs
1
Reasoning cannot be turned off. Adaptive thinking is always on in Opus 5.5; requests sending thinking disabled return a 400 error at every effort level. Because thinking tokens are billed as output and count against max_tokens, workloads that previously ran without thinking can produce more output tokens per request. For strictly latency- or cost-sensitive, high-volume calls where any reasoning overhead is unwanted, a lighter sibling or a model that permits disabled thinking may fit better. Mitigate by running at low effort.
2
Below-average output speed at common settings. Independent Artificial Analysis measurements put Opus 5.5 near 79 tokens per second at medium effort and about 71 at high effort—below the reasoning-model median—with time to first token worsened by thinking time. Max effort reaches roughly 97 tokens per second but with far higher verbosity. For interactive, token-streaming UX where perceived speed matters, test low effort or a faster model; Opus 5.5 is tuned for quality on long tasks, not snappy short responses.
3
Premium pricing tier. Opus 5.5 sits at the top of Anthropic's cost structure. While it is cheaper to run than Opus 5 and benefits from a substantial cache-read discount, its per-token output pricing remains high relative to mid-tier and lightweight models. Independent analysis describes it as somewhat expensive versus models of similar price. For high-volume, cost-sensitive generation that does not need flagship reasoning, Sonnet 5.5 or Haiku-class models are more economical.
4
Launched with restrictive safeguards. Because its biology and cybersecurity capabilities rival Mythos-class models, Opus 5.5 ships with strict guardrails: certain general cybersecurity prompts automatically fall back to Opus 4.8, and higher-tier access requires vetting. Teams building security tooling or sensitive-domain applications should validate that these fallbacks and restrictions do not interfere with legitimate workflows before standardizing on the model.
Best-Fit Workloads
Where this model earns its place
01
Long-horizon agentic coding
Opus 5.5 is Anthropic's most capable Opus model for coding and agents, built to run projects end-to-end. Independent Terminal-Bench 4.0 measurements rise with effort, and Anthropic reports it at default effort beating Opus 5 at max effort for roughly a fifth of the cost. The 1M-token context lets agents hold large codebases in scope, and the cache-read discount lowers cost on repeated-context sessions. Expect below-average token speed, so favor it for quality-sensitive work over snappy interactive edits.02
Financial and document analysis
Anthropic positions Opus 5.5 for financial workflows: reading dense filings, models, and charts, carrying context across an entire deal or reporting cycle, and producing clear summaries of findings. The 1M-token window supports multi-document analysis without truncation, and image input enables chart and table extraction. Its vision and reasoning make it reliable for compliance-sensitive review, though always-on thinking adds output cost per request—tune effort to the depth each document class actually requires.03
Computer use and visual automation
Anthropic describes Opus 5.5 as its best Opus model for vision and computer use, handling multi-step tasks spanning multiple applications that require planning and judgment. It reads screenshots, diagrams, and dense documents at high fidelity, supporting document extraction and interpreting real-world imagery. This suits automation agents that must perceive an interface, plan, and act. Validate the launch safeguards and tool-call behavior for your environment, since forced tool calls are no longer permitted via the API.04
Enterprise knowledge work
Opus 5.5 is pitched as the enterprise daily driver: following instructions precisely, staying in scope, and producing professional-grade spreadsheets, slides, and documents. Anthropic highlights clearer communication that front-loads key information and avoids jargon. For teams standardizing on one high-capability model across research, drafting, and analysis, it reduces the need to route between models—at the trade-off of premium pricing that makes it less ideal for routine, high-volume generation.Pricing in anytokens via AnyAPI
Input
24
₳
Output
120
₳
Cache write
30
₳
Cache read
1.2
₳
Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04