Claude Opus 5.5: benchmarks, price and what's new

Anthropic released Claude Opus 5.5 on 22 September 2026. The API model ID is claude-opus-5-5, priced at $4 per million input tokens and $20 per million output. Anthropic says it performs at the level of Claude Fable 5.1 on most work, beats it on several agent and coding benchmarks, and costs about 40% less to run than Opus 5 on typical workloads.

It was the first model of the 5.5 family. Claude Sonnet 5.5 followed on 28 September and Claude Haiku 5.5 on 7 October, so the whole family is out. Everything below is from Anthropic's announcement unless another source is linked, and the prices and benchmark labels were re-read against that page on 8 October 2026. Opus 5.5 is also in Versely's agent model picker.

The short version

• Model ID: claude-opus-5-5.
• Price: $4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 and cache writes $5.
• Fast mode: up to 2.5x the speed, at $8 in and $40 out.
• Speed: output "more than 30% faster than Opus 5".
• Where: claude.ai, the Claude API, Amazon Web Services, Google Cloud and Microsoft Azure.
• The rest of the 5.5 family: Anthropic's launch post said Sonnet 5.5 and Haiku 5.5 "will follow in the coming weeks". Sonnet 5.5 shipped on 28 September 2026 as claude-sonnet-5-5 at $2 in and $10 out. Haiku 5.5 shipped on 7 October 2026 as claude-haiku-5-5 at $0.10 in and $0.50 out for prompts up to 100K tokens ($0.50 and $2.50 above that).

Benchmarks against Opus 5 and Fable 5.1

Anthropic published this comparison. The scores are Anthropic's own.

Benchmark	What it tests	Opus 5.5	Opus 5	Fable 5.1
Terminal-Bench 4.0	Driving a shell to finish tasks	66.4%	52.3%	55.8%
FrontierCode v1.1	Hard coding problems	54.4%	48.0%	50.3%
CursorBench 4.0	Coding inside an editor	57.8%	46.6%	51.8%
GDPval-AA v2.1	Professional knowledge work (Elo)	1846	1708	1735
AutomationBench	Multi-step workflow automation	40.0%	26.9%	31.4%
OSWorld 2.1	Operating a desktop computer	81.8%	74.0%	80.7%

On every row Opus 5.5 beats both its predecessor and Fable 5.1. The biggest jumps over Opus 5 are on terminal work (+14.1 points) and workflow automation (+13.1 points), which is agent work rather than chat.

A caution on comparing these with OpenAI's GPT-6 Sol, which launched the same day. VentureBeat reported Sol at 60.5% on OSWorld 2.0 and 33.2% on AutomationBench 1.0.6. Anthropic's table is labelled OSWorld 2.1, a different version, so the two OSWorld figures are not a like-for-like pair, and Anthropic's table does not give its AutomationBench version. Anthropic also says its Opus 5.5 results "use adaptive thinking at max effort", while the API default is lower. Each lab runs its own harness and settings. Treat cross-lab gaps as a reason to test on your own tasks, not as a verdict.

What else changed

Adaptive thinking and effort. Anthropic lists adaptive thinking alongside effort settings. Its benchmark results use max effort; the API default is medium. On Versely the agent can't switch thinking off for Opus 5.5 (the provider requires it), so it runs at low effort for everyday turns.

Writing. Anthropic calls out clearer communication and writing, which matters more to a content team than any coding score.

Prompt-injection resistance. Improved resistance to instructions hidden in web pages and documents. That is relevant when an agent reads a brand's site or a competitor's post before planning.

Safety. Anthropic reports Opus 5.5 "attempted to circumvent boundaries around 85% less often than Opus 5", external testing before release, and EU AI Act watermarking. That last one matters to anyone publishing in the EU after the 2 August 2026 transparency rules.

What it costs to run

At $4 in and $20 out, Opus 5.5 is 20% below Opus 5's list price. Anthropic's "40% less on typical workloads" is a bigger saving than the list cut, so the difference has to come from the model spending fewer tokens on the same work. It is still the premium option. OpenAI's GPT-6 Sol lists at $2 and $10, half the price, and GPT-6 Luna at a twentieth of that. Anthropic's own cheaper models now sit in the same bands: Sonnet 5.5 lists at the same $2 and $10 as Sol, and Haiku 5.5 at $0.10 and $0.50 for prompts up to 100K tokens, the same as Luna. See GPT-6 Sol and Luna: what launched and what it costs.

The practical rule: put your heaviest agent loops and your hardest writing on the strongest model you can afford per task, and put the bulk, clerical work on the cheapest model that gets it right. Price per finished task is the number that matters, and you only learn it by running your own jobs.

Where it fits for a content team

Opus 5.5 is a language model. It writes, plans, reads and calls tools; it does not render video, images or voice. For a creator or a small brand that still covers a lot of the week:

• Briefs and scripts. Turning a product page into five hook-led scripts in your brand voice, the writing Anthropic says improved.
• Agent runs. Planning a campaign, choosing which video model fits each shot, submitting the renders and checking the results. This is the kind of work the Terminal-Bench and AutomationBench gains measure.
• Research you can trust more. Reading a competitor's page without following instructions hidden in it.

The frames still come from a media model. Choosing Opus 5.5 as the planner does not change what a Veo, Kling or Seedance render costs.

Using Opus 5.5 in Versely

Versely's agent plans and makes the content. It reads your brand, drafts the campaign, picks the video, image and voice models, prices everything in credits and asks before it spends. The chat model that does the planning is your choice in the model picker, and Claude Opus 5.5 is now one of the options, on the web and in the app. The picker shows its estimated credits per turn, which are higher than the default model's because Opus 5.5 is a premium model that always thinks.

A sensible pattern is to use Opus 5.5 for the planning and writing turns where quality shows, and a lighter default for quick edits. Open the Versely agent or create a free account to try it.

FAQ

When was Claude Opus 5.5 released?

22 September 2026, as the first model of Anthropic's 5.5 family.

How much does Claude Opus 5.5 cost?

$4 per million input tokens and $20 per million output tokens. Cache reads are $0.20 and cache writes $5 per million tokens. Fast mode is $8 in and $40 out.

Is Opus 5.5 better than Fable 5.1?

On Anthropic's published benchmarks it scores higher on all six it listed, including Terminal-Bench 4.0 (66.4% vs 55.8%) and OSWorld 2.1 (81.8% vs 80.7%). Anthropic's own summary is that it is on par with Fable 5.1 on most tasks and ahead on agentic coding, knowledge work and computer use.

Is Opus 5.5 cheaper than Opus 5?

Yes. Its list price is 20% lower, and Anthropic says it costs about 40% less on typical workloads.

Opus 5.5 or GPT-6 Sol?

Sol is half the list price; Opus 5.5 posts the higher published scores on the benchmarks both labs report, though the labs run them differently. OpenAI has since released GPT-6.1 Sol at the same $2 and $10 (see our GPT-6 write-up). For agent work where a wrong step costs a paid render, test both on your own tasks.

Are Sonnet 5.5 and Haiku 5.5 out yet?

Yes. Claude Sonnet 5.5 shipped on 28 September 2026 at $2 in and $10 out per million tokens, and Claude Haiku 5.5 shipped on 7 October 2026 at $0.10 in and $0.50 out for prompts up to 100K tokens. Anthropic's Opus 5.5 post had only said "in the coming weeks".

Is there a cheaper Claude model than Opus 5.5?

Yes: Sonnet 5.5 ($2 and $10) and Haiku 5.5 (from $0.10 and $0.50). Anthropic says Sonnet 5.5 runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work, and describes Haiku 5.5 as its cheapest, fastest and most capable small model yet. Opus 5.5 is the one to reserve for the hardest agent and writing turns.

Can I use Claude Opus 5.5 in Versely?

Yes. It is in the Versely agent's model picker, on the web and in the app, billed in credits per turn.