Key Notes
- Anthropic says Claude Opus 5.5 matches Fable-level performance on most work and costs 40% less to run than Opus 5 on typical workloads.
- Standard API prices are $4 per million input tokens and $20 per million output tokens, a 20% reduction.
- Sonnet and Haiku 5.5 are expected in the coming weeks.
Anthropic released Claude Opus 5.5 on September 22, promising stronger coding and knowledge-work performance at a lower operating cost. The company says the model performs at roughly the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5 on typical workloads.
The release begins a new Claude 5.5 family. Anthropic says Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks, extending improvements across its model range. It has not supplied specific launch dates for those two models.
The immediate commercial question is how much capable AI work a customer can buy, rather than whether every benchmark produces a new record. For developers running agents through lengthy code changes, both the price of each token and the number of attempts needed to finish a task determine the bill.
What the 40% Cost Reduction Includes
Anthropic’s announcement distinguishes the headline running-cost reduction from its API price cut. Standard rates are $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Those changes represent a 20% reduction in each base rate.
Its larger savings estimate also reflects cheaper cached input and lower token consumption. Cached reads cost $0.20 per million tokens, compared with $0.50 previously. Anthropic says output generation is more than 30% faster.
For an illustrative request using one million uncached input tokens and one million output tokens, the base token charge falls from $30 to $24. That calculation excludes caching, tools and other charges. Reaching the larger claimed saving requires the actual workload to benefit from additional efficiencies; it is not a uniform 40% discount on every request or subscription.
Anthropic also offers a faster processing option at $8 per million input tokens and $40 per million output tokens. This makes the trade-off explicit: a team can pay more for latency, or prioritize the lower standard rate when completion time is less important.
The change follows Claude Fable 5.1, which already put price and performance at the center of Anthropic’s product positioning. Opus 5.5 gives customers another model to evaluate before sending every demanding task to the most expensive tier.
Strong Benchmark Results With Important Boundaries
In the launch table, Opus 5.5 leads GPT-5.6 Sol on all six tests with published scores for both models. Against GPT-6 Astra, it leads on four of six overlapping tests. Those counts describe the company’s selected comparison, not every available evaluation.
One standout is Terminal-Bench 4.0, where Anthropic reports 66.4% for Opus 5.5 against 57.9% for Astra. Astra remains ahead in the same release table on AutomationBench and Terminal-Bench-Science. The mixed results matter because coding, business workflows and scientific tasks ask different things of an agent.
The public leaderboard checked for this article did not yet list Opus 5.5. It showed Astra at 58.2% using Codex at maximum effort, a different configuration from the one cited by Anthropic. The launch’s 66.4% result should therefore be treated as company-reported here, rather than independently confirmed by that listing.
Anthropic’s Terminal-Bench comparison uses Opus at xhigh effort and Astra at high effort. Some launch evaluations also allow earlier Claude models to complete work blocked by production safeguards. Such results describe the evaluated system, including its settings and fallback behavior, rather than an isolated model under identical conditions.
For a buyer, counting benchmark wins is less informative than testing the same repository, instructions and acceptance criteria across candidates. A model that produces a correct patch still needs to preserve intended behavior, avoid unrelated changes and leave work that another engineer can review.
Faster Audits Are Promising, but Still Anecdotal
Anthropic cites an early tester who audited and fixed a 200,000-line codebase in under three hours with Opus 5.5. The corresponding Opus 5 run took more than 20 hours and consumed about 2.5 times as many tokens, according to the company.
That is a useful illustration of the potential gain, but it is one reported comparison. Repository structure, test coverage, permissions and the definition of a completed audit can all affect elapsed time. The account does not establish a general sevenfold speedup for software engineering.
For a code review team, a faster first pass could make it practical to inspect more changes or revisit an old backlog. The economic benefit would depend on whether reviewers spend less time correcting the agent and whether the resulting fixes pass meaningful checks. Speed without accepted output can simply move work from the model to the human reviewer.
What Changes for API Developers
The model’s documentation lists a one-million-token context window and a standard maximum output of 128,000 tokens. It accepts text and images and produces text. Developers can access it through the Claude API and supported Amazon, Google and Microsoft cloud platforms.
Large context makes it possible to supply extensive project material, but capacity is not a measure of comprehension. Teams still need to check whether the model finds the relevant dependency, follows the current requirement and distinguishes active code from obsolete examples.
Upgrading existing integrations may require more than changing the model name. Anthropic’s migration guide says adaptive thinking cannot be disabled and forced tool selection is unsupported. The default effort level also changes from high in Opus 5 to medium in Opus 5.5.
There is a visible application change as well: narration between tool calls moves into thinking blocks and can be absent from the default text returned to an interface. Developers who display progress messages need to handle the new format. Otherwise, a working agent may appear silent while it continues its task.
A Broader Claude Upgrade Is Still Coming
Anthropic is increasing five-hour usage limits for Pro, Max, Team and seat-based Enterprise subscribers. The release also carries safeguards for sensitive biology and cybersecurity work. Improved coding capability does not imply unrestricted access to every use case.
The launch follows the integration of Cowork into Claude Chat, putting greater attention on how models perform inside everyday workflows. For users, the relevant improvement is completing a useful job with fewer delays and less supervision.
Sonnet and Haiku 5.5 could extend those gains to cheaper or higher-volume applications when they arrive. For now, Opus 5.5 offers a concrete set of lower rates and new capabilities to test. Its strongest business case will be demonstrated by accepted work per dollar on a customer’s own tasks, rather than a single launch score.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
