Opus 5.5 / 21 Oct 2026 · 22 min read
What is Claude Opus 5.5? Benchmarks, pricing and real limits
Opus 5.5 has strong benchmark results. The useful question is what each test measures—and whether that result helps with the task you actually need to finish.
Claude Opus 5.5 is Anthropic’s AI model for long-running coding and knowledge work. It can work with text and images, reason through a problem and use connected tools to carry out tasks. The model is the engine; Claude’s apps and coding tools provide the environment around it. Anthropic released it on September 22, 2026. Official model documentation.
My Blender experiment is why I’d judge it by the delivered work: Max had the strongest overall finish, while Extra matched the tree reference better, and none made convincing fruit and ice. In my web research comparison, Opus produced useful, readable overviews, but one run drew on a narrow range of sources. Those results make me more interested in which setting and tools suit the job than in declaring one setting the winner.
If you’re choosing an AI model for coding, research or creative production, this guide is for you. It explains the Opus 5.5 benchmarks, API pricing and practical limits, with results for coding, reasoning, research, writing and professional task completion. Sources and leaderboard snapshots were checked on October 16, 2026. The linked Glitch Orchard experiments have their own methods and limitations.
What changed from previous Opus models?
Always-on thinking
The model reasons before answering. Choose how much effort it spends on the job.
Medium by default
Recheck your old configuration before comparing time, price or quality.
Lower standard token rates
Input/output are 20% below Opus 5; your final task bill still depends on usage.
Different tool-call behavior
Apps built for earlier models may need changes to how they request actions.
Quieter progress updates
Work can continue while an app looks quiet. Its progress display may need updating.
Results still need inspection
Tests, original sources and finished files remain the useful acceptance checks.
The API changes come from Anthropic’s migration notes, and the price comparison from its release announcement.
When an app lets a model open files, run commands or operate other software, it can work through a task in steps. That is an agentic workflow. The app supplies the access; the model decides what to do with the tools it has.
Opus 5.5 always uses adaptive thinking, its system for reasoning before answering. Medium replaces High as the default effort level. Anthropic says the same effort names can involve more thinking per turn than on Opus 5, so recheck time, cost and results when you switch models. What changed in Opus 5.5.
Claude Opus 5.5: context window, thinking and tools
A token is a piece of text the model processes; it can be a word or part of one. The context window is the amount of material it can work with at once, including instructions, conversation and supplied content. The output limit is how much it can produce in one response.
An API lets software send requests to the model. The model ID and token prices below are useful if you are building something around Opus. If you use it through an app, that app supplies its own interface, tools and billing arrangements.
The company behind the model.
claude-opus-5-5The identifier developers send to the API.
It reads images; its native answer is text.
Context capacity, not a promise of perfect recall.
Standard output limit, separate from context.
Thinking stays active; effort controls its depth.
The default—not the best setting for every job.
Fresh events still require current sources.
Standard API rates, not a fixed project fee.
These are model specifications, not a promise about every app’s limits. A context window describes capacity; it does not guarantee that every detail will be used correctly. Anthropic’s specifications.
Opus 5.5 is a proprietary model with closed weights. Using it through an API or an app is different from downloading a model to run locally. Artificial Analysis’s model profile.
Developer notes: API migration and platform access
API changes for tool calling, progress and refusals
Forced tool selection with any or a named tool is unsupported; applications need the supported auto or none behavior. Progress text between tool calls moves into thinking blocks and is hidden by default, so an existing interface can appear quiet while work continues. A refusal can arrive in a successful HTTP response with stop_reason: refusal.
Check request handling, progress display and refusal handling before assuming that changing the model ID is the whole migration. Anthropic’s migration notes.
Using Opus 5.5 on the Claude Platform and Google Cloud
The API model is available through the Claude Platform and cloud services including Amazon Bedrock, Google Cloud and Microsoft Foundry. Tool access and settings still depend on the chosen service and your application. A model that supports tool use does not automatically gain access to your files or the live web. Platform availability.
Opus 5.5 benchmarks: what the benchmark scores show
Start with the kind of work you need help with. Each percentage below belongs to its own test: a higher number on a different test does not make it a better result. The paired results show how two evaluators measured the same model under different conditions. A fallback means another model is allowed to take over certain requests when safeguards intervene. A result labelled “with fallback” describes that permitted system.
Technical tasks
About 66 in 100 test tasks completed successfully.
Test settings & source
XHigh means Extra High thinking effort: above High, below Max. Production fallback permitted; see the explanation above.
Read the published resultAbout 60 in 100 test tasks completed successfully.
Test settings & source
Max is the highest listed thinking-effort setting. Production fallback permitted; see the explanation above.
Read the published result- For your work
- Relevant if you want an agent to work through technical tasks using typed commands.
- How to read it
- Different effort and evaluation setups explain why these are separate results. They are not two estimates of your project’s chance of success.
Difficult questions
About 68 in 100 difficult test questions answered correctly.
Test settings & source
External tools were available during this test. Production fallback permitted; see the explanation above.
Read the published resultAbout 61 in 100 difficult test questions answered correctly.
Test settings & source
Max is the highest listed thinking-effort setting. Text questions without external tools. Production fallback permitted; see the explanation above.
Read the published result- For your work
- Relevant if you need help working through difficult academic problems across several subjects.
- How to read it
- Anthropic allowed tools; AA used text questions without tools. Compare the conditions, rather than averaging the scores. Everyday reports still need source checks.
Working in apps
81.8% of the available score, including credit for partly completed tasks.
Test settings & source
Partly completed tasks can earn some of the available points. Production fallback permitted; see the explanation above.
Read the published result- For your work
- Relevant if you want an agent to operate apps and work through computer tasks.
- How to read it
- This is not 82 in 100 workflows finished perfectly. Check the completed files, forms and other results.
Scientific coding
A 66.9% score on scientific programming tasks.
Test settings & source
Max is the highest listed thinking-effort setting. Production fallback permitted; see the explanation above.
Read the published result- For your work
- Relevant if your work involves writing code to solve scientific problems.
- How to read it
- Under review: AA currently flags this evaluation. Keep the launch score provisional.
Anthropic’s three results come from Anthropic’s announcement; Artificial Analysis’s three results come from Artificial Analysis’s release evaluation. Anthropic’s results also allow production fallbacks when safeguards intervene. Tools, effort and evaluation conditions differ between evaluators. The current Artificial Analysis page marks SciCode and CritPt as under review. Anthropic reports a standard error of ±2.6 percentage points for its 66.4% Terminal-Bench result, so small gaps need care.
Opus 5.5 vs Fable, Sol, Luna, Astra and Gemini
These are results from the same evaluator, Artificial Analysis, using its current Intelligence Index v4.3.2 and task evaluations. Start with price versus performance, or switch to broad performance, technical tasks or difficult questions. The settings stay fixed across all four views: Max for six models and High for Gemini. Those names do not imply equal computing budgets. In the price view, each dollar amount is AA’s weighted average model-token cost across its evaluations, including reasoning, caching and failed attempts—not the cost of a successful deliverable. AA cost methodology.
The Intelligence Index uses points, not percentages. It combines ten evaluations and is not a human IQ score. Terminal-Bench measures successful technical tasks; AA’s Humanity’s Last Exam result uses text questions without tools. This is a different setup from Anthropic’s tool-enabled HLE result above. How Artificial Analysis runs these tests.
How does Opus compare?
How much does this level of performance cost?
This pairs a broad test score with AA’s weighted average token cost across those tests. It shows benchmark cost, not the price of your next task or a guaranteed successful result.
Show all 7 models, scores and prices
Scores and costs use each model’s listed configuration
Source: Artificial Analysis · snapshot 16 October 2026 · exact configurations and sources in the expandable list.
Price versus performance: at the settings shown, Opus scores 58 and costs an average $5.98 per AA index task; Sol scores 52 at $0.72. That is six more index points for about 8.31 times the benchmark cost. Luna has the lowest listed average, $0.07, alongside a lower index score of 38.
First decide which models meet your quality requirements, then compare the cost of those models. A cheap result that needs substantial repair can still be expensive to use.
These comparison rows do not supply uncertainty intervals, so small displayed gaps need care.
The three Claude rows allow default fallbacks; these pages do not establish how often another model took over. Gemini’s two task scores are rounded in the source. SciCode and CritPt, which contribute to the broad index, currently carry review notices. Keep this dated snapshot provisional rather than treating it as a permanent ranking. Current AA review status.
The computer-use test, OSWorld, is left out of this comparison because a matching evaluation across all seven models was not verified: the upper Opus card uses OSWorld 2.1 with partial credit, while OpenAI’s reports use a different version and setup.
Agentic coding: Claude Code, coding agents and code review
An agentic coding tool can open project files, change code and run commands to check the result. The model does the reasoning; the surrounding app supplies the tools and access. That environment is often called a harness. Giving the model more thinking time does not automatically give it better tools or make every change correct.
CursorBench 4.0 tests development requests, including editing code, investigating problems, reorganizing code while keeping its behavior, managing jobs and following design requirements. Here is how to read its Opus 5.5 results before choosing an effort setting. CursorBench tasks and results.
What does the percentage mean?
It is Cursor’s benchmark score; higher is better on this evaluation. The published scoring details checked here do not establish that 43.7% means 44 of 100 tasks were completed perfectly. It is not a measure of how much of your code is correct.
What does the dollar amount buy?
It is the average model-token cost per evaluated benchmark task, calculated from input, output and cache prices. The average includes paid attempts that can fail, so it is useful for comparing these test runs rather than quoting a project.
What do Low, High and Max change?
They control reasoning effort: how much work the model is allowed to put into thinking. Max is the highest listed setting. The names are ordered choices, not equal steps in time or computing power, and more effort need not give the best result.
What did the extra thinking buy?
About 76% more cost. Same displayed score: 56.0%.
For 1.8 percentage points more score: 56.0% → 57.8%.
The two charts use different units and scales. Points are observed effort settings; the spaces between them are not equal units of computing power. A task’s cost can include an unsuccessful attempt. Published results.
All five scores and costs
| Opus 5.5 effort | CursorBench score | Avg. token cost per benchmark task |
|---|---|---|
| Low | 43.7% | $1.17 |
| Medium | 52.5% | $2.91 |
| High | 56.0% | $3.97 |
| Extra High | 56.0% | $6.98 |
| Max | 57.8% | $13.43 |
A percentage-point gain is the subtraction: 57.8 minus 56.0 is 1.8 points. These CursorBench observations do not forecast the result or cost of your repository.
These results give a reason to test High on a representative request before paying for Max everywhere. The testing steps below explain how to compare the settings against your own acceptance criteria.
Cursor’s earlier methodology describes real development requests paired with committed solutions and automatic graders, including work involving multi-workspace repositories, production logs and long experiments. Those examples explain the task family. The exact current 4.0 prompts, fixed budgets and conversion from grading to the published percentage were not verified here. Cursor’s methodology.
Reasoning: more effort can produce a lower score
ARC-AGI-2 gives the model examples of coloured input and output grids. It has to work out the rule and apply it to a new grid. That makes this a test of solving an unfamiliar visual puzzle. What an ARC task looks like.
On ARC Prize’s verified ARC-AGI-2 results, Opus 5.5 scores 93.3% at High, 92.5% at XHigh and 91.7% at Max. The ordering runs against the assumption that the highest effort setting must win. For this puzzle test, High leads; the task and setting belong beside the score. ARC Prize results.
Research: search access changes the task
Research adds a different dependency: access to sources. Parallel tests questions that require finding information across sources, answering specialist questions and doing business research. It reports a Search Intelligence Score of 75.4 with search, compared with 44.2 without it. That is a composite score across its research tests. Parallel’s model evaluation.
How Parallel combines its research tests
The score gives equal weight to DeepSearchQA (DSQA), Humanity’s Last Exam and WISER. DSQA uses multi-step research questions and an F1 score, which can give partial credit. The other two use answer accuracy. WISER covers Parallel’s business-research questions; it is not being treated here as a general measure of all web research.
The paired evaluations use the same 100 questions per test with and without Parallel’s search and extraction tools, with fixed tool budgets and no code execution. Reasoning and token limits can vary by model. Parallel’s detailed methodology.
For source-heavy work, check the original publication, the date and whether the conclusion stays inside the evidence. In my web research comparison, Opus added useful topics and readable overviews. One assignment covered astrology; in that run, it opened only GWUP/Skeptix pages. That useful overview still needed a broader source check.
Writing and design: preference needs context
For writing, the question changes again: does the result fit the brief and the voice? ToneBench’s October 13 scorecard gives Opus 5.5 XHigh a Writing Score of 85.8/100 for repeated drafts of ten reference video scripts. That is a judged writing score, not a percentage of sentences that are factually correct. One of the four judges is also an Opus 5.5 model. ToneBench scorecard.
Writing rankings and visual-design results
ToneBench also uses Elo, a relative ranking score. Its leading XHigh and Max Elo ranges overlap, so a small ranking gap does not establish a clear preference between them. This evidence concerns its script-writing task and judging process; it cannot tell you whose newsletter voice you will prefer.
Design Arena lists separate game development, 3D, data visualization, website and SVG results. Those categories help identify what voters preferred. SVG is a format for vector graphics. A pleasing preview still needs its own checks for interaction, accessibility and editable delivery. Design Arena’s Opus page.
The fallback detail that changes the interpretation
The fallback label in the earlier results matters because it can change which model actually completes a request. Vals explicitly describes using Opus 5 and Opus 4.8 as fallbacks. In its September 22 update, SRE Bench falls from 33.59% to 5.34% when fallback-assisted tasks count as failures. The page also contains later live scores, which should be kept separate from that dated update. Vals’ evaluation and fallback disclosure.
For everyday use, that combined service may be what you want. For a model comparison, it matters whether Opus or a fallback finished the work.
Benchmark setup: versions, tools and grading
Reading benchmark scores without losing the setup
Before comparing two models, keep six details beside the result: the benchmark version, model version, effort setting, tools, fallback policy and grading rule. If one is missing, leave it unknown. Giving that gap a confident label makes the comparison look cleaner while telling you less.
For software engineering tasks, look for whether the agent could access the internet, how it was allowed to run regression tests and what counted as a completed task. A benchmark that asks for one patch is a different challenge from one that asks the agent to maintain a project through a sequence of changes.
Also separate a percentage score from a ranking metric such as Elo. A ten-point Elo increase does not mean ten percent more tasks completed. Pass@1 reports performance on a single attempt. Allowing repeated attempts or awarding partial credit changes what the result tells you. The number needs its definition as much as it needs the model name.
Can Opus 5.5 complete professional work reliably?
Good answers to individual questions are only part of a work assignment. The next tests ask whether the agent can work through a multi-step assignment and deliver a result that meets all the requirements. A high relative rank can coexist with a low completion rate on that stricter measure.
On Agents’ Last Exam, Snorkel lists Opus 5.5 at 38.2% overall pass rate using Claude Code at Max effort. The separate score column is 63.2%; those are different metrics and should not be swapped. Agents’ Last Exam leaderboard.
DAYJOB uses file-based professional assignments. Its healthcare examples include finding why a treatment claim was denied and drafting a recommended appeal in Word. A finance example asks the agent to check a fund’s reporting files and write a short memo on whether the reports are ready to send. A task passes only when every required check is met. Opus 5.5 passes 24.7% of healthcare attempts and 23.9% of finance attempts in the reported evaluation. That demanding definition of success makes the gap between a mostly plausible deliverable and a fully correct one visible. DAYJOB paper and task examples.
METR’s predeployment assessment describes an incremental improvement over Fable 5.1 in AI R&D capability, with continuing weaknesses on difficult, extended tasks. It does not support treating Opus as a fully autonomous researcher. METR also explains its access and Anthropic’s review of the published summary. METR assessment.
Resisting a bad argument is only half the job
Another part of working with an assistant is correcting it. Big-Big-SycoBench tests whether a model changes its answer when presented with a valid or invalid argument. It adds a less comfortable result. In its tested rescoring setup, Opus 5.5 Max followed 0.8% of invalid arguments—but also only 32.2% of valid arguments. The run used a stratified subset of 100 artifacts, rather than all 450 in the dataset. The authors distinguish resistance to bad reasoning from willingness to accept a good correction. Big-Big-SycoBench paper.
For a working assistant, both matter. You want it to challenge a mistaken premise and change its answer when the evidence warrants it. Include this in your own evaluation: can the model tell a useful correction from pressure to agree?
What my Opus 5.5 Blender test adds
My Claude Blender review takes this from published scores to delivered files. It compares thirteen tasks across five runs labelled Low, Medium, High, Extra and Max. Each label has its own set of results in the linked review.
One task asked for an editable 3D reconstruction of an oak-tree-and-swing reference image. Another asked for a finished 15-second energy-drink commercial, including close-ups of the can, fruit, ice and a moving splash. The test therefore had to consider whether the scene could be edited, whether it followed the reference and whether the finished animation looked convincing.
Max had the strongest overall finish in that batch. Extra matched the tree reference better. The commercial’s fruit and ice still fell short of convincing photorealism across all five runs. A functioning file and a convincing image are separate accomplishments. The full review preserves the outputs and revisions so you can inspect those differences.
What this Blender comparison can establish
The review explains shared contexts, iterations and incomplete comparable cost records. It adds concrete examples without establishing a controlled causal ranking of effort levels. The five run labels stay as recorded; “Extra” is not being silently relabelled as a different API setting.
How much does Claude Opus 5.5 cost?
The earlier graphs show what evaluation runs cost. This section explains the rates used when software calls Opus directly through its API. Standard list prices are $4 per million input tokens and $20 per million output tokens. Input is what you send; output is what the model produces, including billed reasoning. Cache reads reuse previously processed input and cost $0.20 per million tokens. Fast mode has a separate price. A Claude subscription uses a separate billing arrangement. Anthropic’s release pricing.
Those standard input and output rates are 20% below Opus 5’s $5/$25. Cache reads fell by 60%, from $0.50 to $0.20. Anthropic’s separate claim of roughly 40% lower task costs concerns its typical workloads at default settings; it is not the size of the base token-price cut. Its reported output-generation speed improvement is over 30%. Anthropic’s cost and speed comparison.
Comparing Claude API prices: Opus 5.5, Sonnet 5.5 and Claude Fable 5.1
These are standard token rates, rather than the task averages in the benchmark graphs. Opus costs $4/$20, Sonnet $2/$10 and Fable $10/$50 per million input/output tokens respectively. Current model comparison and prices.
How Anthropic positions these Claude models
Anthropic positions Opus for extended coding and knowledge work, Sonnet for a combination of speed and intelligence, and Fable for demanding reasoning or work where Opus at higher effort still falls short. That describes the provider’s intended uses; the task-specific results earlier in this article show why the name alone cannot choose a model for you.
API pricing details: tokens versus a completed task
A simple uncached request with 100,000 input tokens and 10,000 output tokens costs $0.60 in model-token charges: $0.40 for input plus $0.20 for output. For an accepted deliverable, also count additional thinking, retries, tool charges and review time.
Try the token calculator: two illustrative 10,000-row examples
These presets use standard uncached Opus rates and one call. Each assumes a size for every input row and adds 10,000 output tokens. They are arithmetic examples rather than measured database runs.
Estimate the token bill
Enter estimated billable tokens, including any reasoning tokens in the output estimate. This calculator does not estimate tokens from your database.
$0.40 input + $0.20 output. Illustrative model-token subtotal, not a measured database job.
USD, standard uncached rates per million input/output tokens. A million-token input leaves no room for instructions and output in a nominal 1M context, so real work may need splitting. Price source.
Assumptions behind the two database examples
| Assumed workload | Input charge | Output charge | Model-token subtotal |
|---|---|---|---|
| 10,000 rows × 10 input tokens, plus 10,000 output tokens | $0.40 | $0.20 | $0.60 |
| 10,000 rows × 100 input tokens, plus 10,000 output tokens | $4.00 | $0.20 | $4.20 |
The presets give 100,000 or one million input tokens. A million-token input already fills the nominal context allowance before instructions, schema and output, so real work may need splitting. Repeated calls, reasoning tokens, caching, tools, database hosting and human review change the bill.
CursorBench cost comparison: four Claude models
This chart compares measured average costs for CursorBench 4.0 development tasks across four Claude models. Switch between cost and score to see what extra spending bought in this test. Each line follows five ordered effort settings; the spacing between them does not represent equal computing budgets. This is a separate evaluation from Artificial Analysis’s broader index above. CursorBench source and task description.
CursorBench costs by model and effort
Average token cost for Cursor’s development tasks. This is a different evaluation from the Artificial Analysis comparison. Switch to scores to compare the results.
Same effort names do not mean equal compute. Costs are averages per evaluated benchmark task, not prices per successful completion. CursorBench 4.0, checked 16 October 2026; Sonnet price update 7 October.
See all 20 scores and costs
| Model | Effort | CursorBench score | Avg. cost per benchmark task |
|---|---|---|---|
| Claude Opus 5.5 | Low | 43.7% | $1.17 |
| Claude Opus 5.5 | Medium | 52.5% | $2.91 |
| Claude Opus 5.5 | High | 56.0% | $3.97 |
| Claude Opus 5.5 | Extra High | 56.0% | $6.98 |
| Claude Opus 5.5 | Max | 57.8% | $13.43 |
| Claude Sonnet 5.5 | Low | 35.8% | $0.37 |
| Claude Sonnet 5.5 | Medium | 39.2% | $0.52 |
| Claude Sonnet 5.5 | High | 47.8% | $1.20 |
| Claude Sonnet 5.5 | Extra High | 53.1% | $2.81 |
| Claude Sonnet 5.5 | Max | 55.5% | $7.05 |
| Claude Fable 5.1 | Low | 45.1% | $5.44 |
| Claude Fable 5.1 | Medium | 46.8% | $7.05 |
| Claude Fable 5.1 | High | 49.2% | $9.08 |
| Claude Fable 5.1 | Extra High | 51.6% | $13.01 |
| Claude Fable 5.1 | Max | 51.8% | $17.28 |
| Claude Opus 5 | Low | 40.7% | $4.87 |
| Claude Opus 5 | Medium | 43.3% | $6.94 |
| Claude Opus 5 | High | 44.7% | $9.00 |
| Claude Opus 5 | Extra High | 46.1% | $11.43 |
| Claude Opus 5 | Max | 46.6% | $11.95 |
Choosing the best model and effort setting for your task
Start by writing down what would make the result acceptable. For a research report, that might mean every central claim has an opened original source and the dates fit the requested period. For a coding task, it might mean the regression tests pass, the requested behavior works and the change introduces no known security problem. For a scene, it means the saved file opens and the result matches the brief from the required views.
Run the same task with the models and settings you want to compare, using the same tools and starting material. Keep the first outputs as well as revisions. Record which attempts needed another prompt, a manual fix or a restart. Repeat enough times to see whether a small difference survives a new run.
That gives the effort setting a practical meaning: how often it reaches your standard, how long it takes and what you pay for the successful result. It also stops the prettiest single output from quietly becoming the entire benchmark.
Quick answers about Opus 5.5
Is Opus 5.5 the same thing as Claude Code?
No. Opus 5.5 is a model. Claude Code is an environment that can use a model to work with code and tools. Benchmark reports should name both.
Does Opus 5.5 generate images and videos?
Its documented output modality is text. It can write code or direct connected software that produces visual assets. My Blender experiment tests that connected workflow, rather than native video generation.
Is Max the best Opus 5.5 setting?
It depends on the task. CursorBench shows a higher score with a higher listed cost at Max; ARC-AGI-2 gives High the better result. Measure the setting against your acceptance criteria.
Are all public Opus 5.5 benchmarks directly comparable?
No. Versions, tools, harnesses, fallback policies and grading differ. A useful result keeps those conditions beside the score.
If you want to see what that looks like beyond a leaderboard, open the complete Blender task comparisons or the original-source checks in my research experiment. The outputs make it easier to decide what you would actually accept.