Anthropic ships Claude Opus 5.5: the same level of work at 40% lower cost
Anthropic announced Claude Opus 5.5 on September 22, 2026. The company says the model performs at the level of Claude Fable 5.1 on most work and does it for 40 percent less than Opus 5. Input and output now cost 4 and 20 dollars per million tokens, cache reads 0.20 dollars, and output generation is more than 30 percent faster than Opus 5.
A release roughly two months after Opus 5
Anthropic released Claude Opus 5.5 on September 22, 2026. The headline claim sits on the cost side rather than the capability side: the company says the model works at the level of Claude Fable 5.1 on most tasks while costing 40 percent less to run than Opus 5. TechCrunch notes the release came about two months after Opus 5. The model is callable on the Claude Platform under the id claude-opus-5-5, and Anthropic says it is available across all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
Speed is the second axis of the announcement. According to Anthropic, Opus 5.5 generates output more than 30 percent faster than Opus 5 and offers a fast mode in Claude Code and the Claude Platform that reaches up to 2.5 times the speed, priced at 8 dollars per million input tokens and 40 dollars per million output tokens. TechCrunch adds two observations about how the model writes: it reaches for jargon less often and puts the important information first.
Three price lines down, benchmark scores up
Pricing moved on three separate lines. Input tokens are 4 dollars per million and output tokens 20 dollars, against 5 and 25 dollars on Opus 5, a drop of 20 percent. In percentage terms the sharpest cut is in caching: cache reads fall from 0.50 dollars to 0.20 dollars per million tokens, a difference of 60 percent. Cache writes are listed at 5 dollars per million tokens.
Anthropic also published comparative benchmark results. On Terminal-Bench 4.0, Opus 5.5 scored 66.4 percent against 52.3 percent for Opus 5. On GDPval-AA v2.1 the Elo rating is 1846 against 1708. FrontierCode v1.1 reads 54.4 percent against 48.0 percent, CursorBench 4.0 57.8 percent against 46.6 percent, and OSWorld 2.0 81.8 percent partial against 74.0 percent. Every one of those figures comes from Anthropic's own measurements; neither source carries an independent verification.
Cheaper tokens, faster output and a behavioral audit
The third theme Anthropic puts forward is its automated behavioral audit. On that audit the company calls Opus 5.5 "the strongest-performing model we've tested to date". That assessment also belongs to Anthropic's own process, and the sources contain no third-party review of it. A model getting cheaper and faster in the same release, as the announcement describes it, produces a picture that teams choosing by price list can read directly.
Cost falling on three axes at once while reported benchmark scores rise is the center of gravity here, and the detail worth noting is that the largest cut in percentage terms landed on cache reads. We looked at how agentic workloads are moving on the platform side in our September 11 piece on the OpenAI Agents API. What the sources leave out deserves a note too: the date the new prices take effect, how long Opus 5 remains on offer and the context window are absent from both reports.
The sources never mention Türkiye, so this section is commentary
Neither source contains a single statement about Türkiye: nothing on the conditions of access from the country, no price in Turkish lira, no measurement of performance on Turkish input and output, and no reference to local regulation. Anthropic's announcement is product and price focused, and the TechCrunch report comes out of a US publishing line. What follows is therefore not a fact drawn from the sources but explicitly UNALSOFT commentary.
Our reading is this: the published figures are in dollars, so the arithmetic for a team in Türkiye changes in that same currency, and everything beyond that is simply missing. The unknowns should be listed one by one. The sources do not say: the conditions of access from Türkiye, how performance on Turkish is measured, which cloud provider regions carry the model, where data is processed, how KVKK would treat it, the effective date of the new prices, how long Opus 5 stays available, the context window, whether the benchmarks have been reproduced independently, or any figure on enterprise use in Türkiye. None of those questions is answered in either report.
The UNALSOFT view
For us this is a budget question rather than a model recommendation. The cost of an AI workflow is not one number but the sum of input, output and cache reads, and because the cut Anthropic announced applies to all three lines, a fair comparison has to look at each line separately. At the same time every published measurement is the vendor's own test, so the decision belongs to a trial run against your own workload. That is why our agentic AI work keeps the model choice replaceable instead of fixed: if roughly two months separated Opus 5 from Opus 5.5, not pinning a workflow to a single model id is a defensible position. Because the sources say nothing about Türkiye, this article makes no prediction about access or local use.
Sources
Anthropic, Claude Opus 5.5 announcement · TechCrunch, Anthropic releases Opus 5.5 with lower prices
Can you recalculate what your AI workflows cost the day the model changes?
A short conversation is enough to build your agentic workflows without locking them to one model and to see the cost line by line.