Claude Opus 5.5: Benchmarks, Pricing, and Real-World Performance
Anthropic's Opus 5.5 explained: terminal bench and GDPval scores, 40% lower cost per task, and how it performs in hands-on coding tests.
What is Claude Opus 5.5?
Claude Opus 5.5 is Anthropic’s newest flagship model, the first release in its “Claude 5.5” family. It matches or beats GPT-5.1 (referred to in early commentary as “Fable 5.1”) on most agentic and coding benchmarks while costing roughly 40% less to run per task than its predecessor, Opus 5. Anthropic paired the release with a smaller price cut on raw API tokens, faster output generation, and a rewritten communication style meant to sound less like typical AI-generated text.
TL;DR
- Terminal-bench 4.0, one of the least saturated agentic coding benchmarks, is where Opus 5.5 shows its biggest jump, scoring 66.4% versus 57.9% for the next closest competitor and roughly 10 points ahead of GPT-5.1.
- GDPval 2.1, OpenAI’s real-world knowledge-work benchmark covering tasks like spreadsheets, documents, and email, shows Opus 5.5 with an ELO score around 1846, more than 300 points above GPT-5’s Astra variant.
- Cost per task, not cost per token, is the real story: Anthropic cut list pricing by about 20% ($4/$20 per million input/output tokens versus $5/$25 for Opus 5), but the model also needs fewer tokens to finish the same work, producing the claimed 40% total savings.
- Reasoning effort settings matter a lot: on some benchmarks medium effort outperforms high or max settings, meaning users should test effort levels rather than assume “max” is always best.
- Writing style got a deliberate overhaul, with shorter, more direct explanations that put key information first, aimed at reducing the “AI slop” feel of earlier Claude outputs and making it easier to track multiple parallel agent sessions.
- Hands-on demo testing (3D visualizations, real-time ISS tracking, autonomous research-to-website generation) showed strong visual and agentic capability, though not every output impressed on first look.
- Anthropic also improved safety scores, reporting the best results yet on its internal behavioral audit, while adding stricter safeguards for biology and cybersecurity given the model’s increased capability in those domains.
How does Opus 5.5 perform on benchmarks?
Opus 5.5 posts state-of-the-art or near-state-of-the-art results across most agentic benchmarks tracked at launch. On Terminal-Bench 4.0, which measures a model’s ability to execute real terminal commands as part of agentic coding work, it scored 66.4%, ahead of GPT-5’s Astra at 57.9% and GPT-5.1 at 55.8%. That’s a meaningful jump over Opus 5’s 52.3%.
On Frontier Code v1.1, an agentic coding benchmark, Opus 5.5 scored 54.4% versus 50% for GPT-5.1 and 53.3% for Astra, a tighter race. On Cursor’s own benchmark, it hit 57.8%, ahead of GPT-5.1’s 51%. GDPval 2.1, which scores models on real knowledge-work tasks using an ELO-style system, is where the gap widens most dramatically: Opus 5.5 hit roughly 1846 versus GPT-5.1’s 1735 and Astra’s 1542.
Not every benchmark favored Opus 5.5. On Automation Bench it landed slightly behind Astra (40% versus 41.4%), and on Terminal-Bench Science it came in second to Astra as well. Computer-use and visual chart recognition benchmarks showed only marginal gains over GPT-5.1. The Frontier Code benchmark also produced an unusual pattern where medium reasoning effort outperformed high and extra-high settings, a result that likely reflects quirks in that benchmark’s task distribution rather than a general flaw in the model.
Why does cost per task matter more than price per token?
Anthropic cut Opus 5.5’s list price to $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads dropped to 20 cents from 50 cents, and cache writes to $5 from $6.25. That’s roughly a 20% reduction in raw pricing.
But Anthropic also claims a 40% total cost reduction on typical workloads, which only makes sense if the model is also more efficient per task, not just cheaper per token. A model priced lower but requiring far more tokens to reach an answer can still end up costing more in practice. Anthropic states Opus 5.5 needs less compute to serve than Opus 5, and independent benchmark plots (plotting quality against cost per task) show it sitting in the “high quality, low cost” quadrant across several evaluations, including GDPval and Automation Bench. On some coding benchmarks, medium reasoning effort finishes tasks for under a dollar while scoring higher than max effort runs that cost several dollars per task, underscoring that reasoning-effort tuning is now a real cost lever for developers.
What changed in how Opus 5.5 writes and communicates?
Anthropic says it made deliberate changes to the model’s communication style after feedback on Opus 5’s writing, which many users described as stiff or unnatural. Opus 5.5 is designed to lead with the most important information, use plainer sentence structure, and follow explicit formatting instructions more consistently.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
In practice, this shows up in shorter, more direct explanations of what a coding agent just did, rather than long hedged preambles. This matters most for people running multiple agent sessions in parallel, since concise, front-loaded summaries make it faster to switch between threads and pick up context. The model is generally positioned as strongest for coding and knowledge work rather than creative writing, where other models are still seen as producing more natural prose.
How did Opus 5.5 perform in hands-on testing?
Beyond benchmarks, demo testing covered a range of generative and agentic tasks. A black hole simulation prompt produced a visually detailed result with adjustable parameters like density, though brightness needed manual tuning. A real-time International Space Station tracker combined live API calls with accurate positioning and small touches like correctly placing the sun’s position, which held up well as a test of both visual output and live data handling.
An autonomous “research and build a product launch page” prompt produced mixed initial impressions. The visual polish wasn’t uniformly strong, but the model added working animated benchmark comparisons with accurate numbers, including a self-generated comparison between Opus 5 and Opus 5.5. Reasoning effort was set to medium by default in these tests, with Anthropic recommending higher settings for more complex builds. Other tests involving Three.js scenes and detailed voxel art style outputs followed prompts closely and produced some of the more detailed generative visuals seen from a Claude model to date.
Is Opus 5.5 worth switching to?
For teams already running agentic coding workflows, the terminal-bench and GDPval gains are large enough to matter, and the lower cost-per-task profile makes it attractive even before factoring in the modest token price cut. Anthropic also raised the 5-hour usage limit and added the ability for users to reset their own rate limits, a feature likely influenced by competitive pressure from OpenAI.
The tradeoffs are narrower than they might first appear: Opus 5.5 doesn’t win every benchmark (Astra still edges it out on Automation Bench and Terminal-Bench Science), and reasoning-effort settings need tuning per task type rather than defaulting to the highest option. Anthropic also reports improved safety audit scores alongside additional safeguards for biology and cybersecurity use cases, given the model’s increased capability in dual-use domains, which may affect what unrestricted users can query without joining a verification program.
Frequently Asked Questions
What is the difference between Opus 5.5 and Opus 5?
Opus 5.5 scores higher on most agentic and coding benchmarks, costs about 40% less per task on typical workloads, generates output over 30% faster, and uses a more direct, less “AI-sounding” communication style.
How much does Opus 5.5 cost through the API?
List pricing is $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cache reads and writes were also reduced.
Is Opus 5.5 better than GPT-5.1?
It outperforms GPT-5.1 on most benchmarks shown at launch, including terminal-bench and GDPval, though the two are close on some agentic coding benchmarks like Frontier Code.
Where can I access Opus 5.5?
It’s available through Claude Code, the Claude desktop app (after updating), and the API using the Opus 5.5 model tag.
Does reasoning effort level affect cost and quality?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Yes. On several benchmarks, medium reasoning effort produced comparable or better scores than high or max settings at a fraction of the cost per task, so testing different effort levels for your specific workload is worthwhile.
