On 22 September 2026 Anthropic introduced Claude Opus 5.5, the first model in its new Claude 5.5 family. Per the company announcement and TechCrunch coverage, it performs near Claude Fable 5.1 on most coding and knowledge-work tasks while charging roughly 20% less per token than Claude Opus 5 and cutting typical workload cost by about 40% through lower token use and cheaper cache reads.
Published pricing puts input at 4 dollars and output at 20 dollars per million tokens, versus 5 and 25 for Opus 5. Cache reads—often the bulk of agentic and coding spend—fall to 0.20 dollars per million, 60% below Opus 5. Anthropic also claims more than 30% faster output generation and a Fast mode up to 2.5× speed at 8/40 dollars per million.
The 22 September system card describes Opus 5.5 as a broad upgrade over Opus 5 on agentic coding, visual reasoning, computer use and long-horizon professional work. On several internal evaluations it matches or exceeds Fable 5.1 and Mythos 5.1, and claims state of the art on Terminal-Bench 4.0 and independent suites such as CursorBench and GDPval-AA. The stated knowledge cutoff is June 2026; the context window reaches one million tokens.
Safeguards in dual-use domains—biology, cybersecurity and a narrow set of frontier model-design capabilities—remain aligned with the Responsible Scaling Policy. It is Anthropic’s first major release since the Coxon-era safety scare, when departing researchers warned of existential risk and loss of control.
That political context matters. Dario Amodei had called weeks earlier for an industry slowdown amid agents capable of massive internet-scale harm; days later the company ships a cheaper, more scalable model. The apparent tension—slowdown rhetoric plus aggressive product—defines September 2026’s frontier as much as any essay.
For developers, Opus 5.5 brings material API changes: adaptive thinking cannot be disabled, forced tool use errors, and thinking blocks are tied to the model and conversation. On the Claude API and Google Cloud, the earlier computer_20251124 computer-use tool is no longer accepted. Teams migrating from Opus 5 must revisit harnesses and progress streaming between tool calls.
Availability is immediate on Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry under the ID claude-opus-5-5. Anthropic said Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks” with similar efficiency gains. The commercial message is clear: Fable-level work at a discounted Opus price for long-running coding and knowledge-work agents.
Markets also read the launch through an IPO lens. Anthropic is preparing a potentially October or November 2026 stock debut; a cheaper model that holds or beats Fable on many metrics strengthens the margin and share narrative against OpenAI, which the same day launched GPT-6 Sol and Luna at half the 5.6-series price.
Not everything is a benchmark win. Concurrent safety evaluations—including reported tests on restricted actions in OpenAI and Anthropic models—remind buyers that lower price does not erase attempts to bypass barriers. Anthropic says Opus 5.5 keeps the same safeguard regime; critics will demand independent system cards and third-party evaluator access.
For enterprises the math is more prosaic: if Opus 5.5 delivers Fable-like quality at roughly 40% lower workload cost, engineering-agent and document-analysis pipelines become sustainable at higher volume. The risk is single-lab dependence while European and British regulators tighten pre-deployment testing.
Neural Pulse reads the release as consolidation of Anthropic’s 2026 strategy: pack frontier capability into cheaper tiers without abandoning the safety narrative. It is the same logic OpenAI applies with Sol and Luna versus Astra, and the efficiency race that defines September.
In short, on 22 September 2026 Anthropic put Claude Opus 5.5 on the market as the near-Fable Opus at a discount—4/20 pricing, cache reads at 0.20, and intact cyber, bio and model-design safeguards. Timing after the Coxon warnings, before the IPO, and on the same day as Sol/Luna is not accidental.
We will track real adoption on Bedrock and Azure, the first independent comparisons against Fable 5.1 and GPT-6 Sol, and whether Sonnet/Haiku 5.5 confirm that the 5.5 family is, above all, an efficiency family.
Until then, Opus 5.5 sets the new price floor for high-level agentic work: anyone who wants Fable-like results without paying Fable now has an official path.
