Our Blog

Blog Index 

Claude Opus 5.5: Anthropic's New Flagship Matches Fable-Level AI at 40% Lower Cost

Posted on 25th Sep 2026 06:04:37 in Artificial Intelligence, Machine Learning

Tagged as: Anthropic, Claude Opus 5.5, AI Models, Fable 5.1, Frontier AI

Anthropic has released Claude Opus 5.5, the first model in its new Claude 5.5 family, and the company is making a striking claim about the economics of frontier AI. According to Anthropic, the new model performs at the level of Claude Fable 5.1 - its most capable widely released model - on most work, while costing about 40 percent less to run than Claude Opus 5, the flagship it replaces just two months after launch.

The release landed on September 22, 2026, and it says as much about the state of the AI race as it does about the model itself. Within ninety minutes of Anthropic's announcement, OpenAI released its own updated models, GPT-6 Sol and GPT-6 Luna, at half the prices of their predecessors. Frontier AI is suddenly competing on cost as aggressively as on capability, and businesses that build on these models now have to think about both.

Claude Opus 5.5 is available immediately on the Claude API under the model identifier claude-opus-5-5, as well as on Amazon Bedrock, Amazon Web Services, Google Cloud, and Microsoft Foundry. Consumer and team users get it through Claude's Pro, Max, Team, and seat-based Enterprise plans. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks with many of the same improvements to performance, efficiency, and safety.

Fable-Level Performance at a Lower Price

Anthropic positions Opus 5.5 as a major step up from Opus 5 and a near-peer of Fable 5.1, the top-tier model it released on September 1. On the company's published benchmarks, Opus 5.5 leads in agentic coding, computer use, and knowledge work:

  • Terminal-Bench 4.0 (agentic coding): Opus 5.5 scores 66.4 percent, ahead of Fable 5.1 at 55.8 percent, GPT-6 Astra at 57.9 percent, Opus 5 at 52.3 percent, and GPT-5.6 Sol at 37.3 percent.
  • CursorBench 4.0 (coding agents on real multi-file tasks): 57.8 percent, versus 51.8 percent for Fable 5.1 and 41.7 percent for GPT-5.6 Sol.
  • FrontierCode v1.1: 54.4 percent, beating GPT-6 Astra's top score of 53.3 percent at roughly a fifth of the cost per task, according to Anthropic.
  • GDPval-AA v2.1 (knowledge work): 1,846, compared with 1,735 for Fable 5.1 and 1,708 for Opus 5.
  • Humanity's Last Exam (with tools): 67.7 percent, and OSWorld 2.0 (computer use): 81.8 percent partial.

Benchmark tables only tell part of the story, and Anthropic itself notes that at this level of capability the margins between models have become a less reliable guide to real-world differences. The more interesting evidence comes from early testers. One tester completed a 680,000-line code migration in less than a day, work Anthropic says would have taken an engineering team weeks. A 200,000-line codebase audit that took Opus 5 more than 20 hours was finished by Opus 5.5 in under three hours, using about 2.5 times fewer tokens. In an internal test, both Opus 5.5 and Fable 5.1 translated HAProxy from C into Rust and passed nearly all of the software's own regression tests, but Opus 5.5 finished in 9.5 hours against 12 hours for Fable 5.1 - and cost 51 percent less.

Enterprise testers report similar gains. Quantium said a complex coding task that previously took 38 prompts over four days came in at 11 prompts over three hours. Clio let the model run unattended on a task spanning six code repositories for more than 18 hours and reported it stayed on task throughout. Spotify said Opus 5.5 matched Opus 5's quality in about half the turns, time, and output tokens, cutting the cost of that workload by 40 to 50 percent. These are vendor-published testimonials rather than independent audits, so the honest reading is that Opus 5.5 is at least a solid incremental gain with unusually strong efficiency claims attached.

The Pricing Math: Cheaper Tokens, Faster Output

The headline numbers are straightforward. Input tokens cost $4 per million, down from $5 on Opus 5, and output tokens cost $20 per million, down from $25. Cache reads - which dominate the cost of long agentic and coding sessions - fall 60 percent, from $0.50 to $0.20 per million tokens. Cache writes drop from $6.25 to $5 per million. A Fast mode priced at $8 per million input tokens and $40 per million output tokens runs up to 2.5 times the standard speed, as a research preview on the Claude API.

Anthropic says the combination of lower unit prices and fewer tokens spent per completed task nets out to a 40 percent cost reduction on typical workloads at default settings, and that Opus 5.5 generates output more than 30 percent faster than Opus 5. The company is also raising five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans and giving subscribers a rate-limit reset they can save and use whenever they choose.

There is a catch worth flagging for anyone running a cost model. Opus 5.5 changes two defaults: thinking can no longer be switched off, and the default effort level is now medium where Opus 5 ran at high. Anthropic's own documentation notes the model tends to think more per turn at a given effort level, so teams that carry old settings over unchanged may not see the 40 percent saving their spreadsheet predicts. The migration guide recommends a fresh effort sweep and a canary rollout before moving production traffic.

Safety: The Strongest Audit Scores Anthropic Has Recorded

Opus 5.5 arrives with the strongest safety story Anthropic has attached to a launch to date. On the company's automated behavioral audit - an alignment evaluation that runs the model through thousands of simulated scenarios - Anthropic says Opus 5.5 is the best-performing model it has tested. In a new evaluation measuring how often a model tries to work around the boundaries it has been given, Opus 5.5 attempted to bypass containment about 85 percent less often than Opus 5 or Claude Mythos 5.1, and every recorded attempt was low severity and flagged by the model itself.

Because the model is comparable to Mythos 5.1 in biology and cybersecurity, it is the first Opus model to ship with safeguards similar to Fable 5.1's across cybersecurity, biology, and distillation. Requests that trigger those safeguards fall back transparently to earlier models - cybersecurity tasks route to Opus 4.8, while biology and frontier model development tasks route to Opus 5. Vetted organizations can apply to Anthropic's Life Sciences Verification Program for biology research access, and the company is expanding its Cyber Verification Program to cybersecurity practitioners in the coming weeks.

The launch is also Anthropic's first release since CEO Dario Amodei published an essay calling for the industry to pace the frontier - deliberately slowing capability gains so risk prevention can keep up. The company paired that message with a commitment to give third-party evaluators permanent, employee-level access to verify its safety practices. External organizations including Frontier Design and METR tested Opus 5.5 before release. The model ships with zero data retention options, watermarking measures to comply with the EU AI Act, and the preserved-thinking anti-distillation safeguard introduced with Fable 5.1 for API accounts created on or after August 31, 2026.

What Developers Need to Know Before Switching

Upgrading from Opus 5 is not just a model-string change. Anthropic documents four breaking changes that will return errors or silently alter behavior for existing integrations:

  • Thinking cannot be disabled. Requests that set thinking to disabled or use a manual token budget return a 400 error; depth is now controlled through the effort parameter instead.
  • Forced tool use is rejected. Tool-choice settings of "any" or "tool" return a 400, so integrations must rely on automatic choice, strict schema validation, or structured outputs.
  • Thinking blocks are bound to the model and the conversation. Opus 5.5 reads reasoning blocks from earlier Opus, Sonnet, and Haiku models but not from Fable or Mythos, and editing earlier turns mid-conversation invalidates them - newer API accounts get a 400 by default.
  • The older computer-use tool is no longer accepted on the Claude API or Google Cloud. Developers must declare the newer toolset and update their agent loops; the older tool keeps working on Amazon Bedrock.

The model keeps the one-million-token context window and 128,000-token maximum output introduced with Opus 5, with a knowledge cutoff of June 2026. Anthropic's recommended starting point for most workloads is now Opus 5.5, escalating to Fable 5.1 only when higher-effort Opus evaluations still fall short, while Sonnet 5 remains the default for high-volume, latency-sensitive work. Opus 5 stays active with a published retirement floor of July 24, 2027, and Opus 5.5's own retirement floor is September 22, 2027 - useful dates for any long-lived production integration.

A Price War in Frontier AI

Context matters here: Anthropic published Opus 5.5 at 16:31 UTC on September 22, and roughly ninety minutes later OpenAI released GPT-6 Sol and GPT-6 Luna at half the prices of their GPT-5.6 predecessors. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, while Luna - aimed at high-volume tasks like document summarization and data extraction - costs $0.10 and $0.50. OpenAI explicitly framed the cuts as savings from better caching and inference being passed on to customers.

For Indian businesses and small teams building on these models, the practical effect is that serious agentic AI keeps getting cheaper. Tasks that were priced out of reach a few months ago - automated code maintenance, document-heavy back-office workflows, multi-step research agents - are edging toward affordability for small companies, not just enterprises. The competitive dynamic cuts both ways, though: with frontier labs now launching price cuts within minutes of each other, procurement decisions based on this month's pricing may look very different by December.

The other lesson from this launch is that capability and safety messaging are now packaged together. Opus 5.5's strongest selling points beyond price are its efficiency on long-running agentic work and its audit scores against misaligned behavior. For a business deciding which model to build on, that combination - cost per completed task, not cost per token, plus verifiable safety behavior - is where the evaluation should start.

Sources

Recent Posts

Catgories

Archives

whatsapp me