Claude Opus 5.5: Fable-Level Performance, Cheaper Than Opus 5

Exactly 60 days after announcing Opus 5, Anthropic released Claude Opus 5.5 on September 22, 2026. At first glance it looks like an ordinary point release. The numbers tell a different story: Opus 5.5 outperforms Claude Fable 5.1, a model 2.5 times more expensive, on most benchmarks, and it does so at a lower price than Opus 5.

For Opus 5, we said "same price, a generational leap." For Opus 5.5, the summary is shorter: pay less, get more.

Summary: Claude Opus 5.5 in 30 seconds

• Release date: September 22, 2026. Available the same day on Claude Platform, AWS, Google Cloud and Microsoft Azure.

• Price: $4 input / $20 output per million tokens. That is a 20% cut compared to Opus 5; the cut on cache reads is 60%.

• Context window: 1 million tokens of input, 128K tokens of output. Knowledge cutoff is June 2026.

• Performance: Beats both Opus 5 and Fable 5.1 on Terminal-Bench 4.0, CursorBench 4.0 and GDPval-AA v2.1.

• Speed: According to Anthropic, it generates output more than 30% faster than Opus 5; fast mode allows up to 2.5x speed.

What happened in the last two months?

The summer months were busy for the Claude family. A quick recap:

• July 24, 2026: Claude Opus 5 was released. At the same price ($5 / $25), it went well beyond Opus 4.8.

• September 1, 2026: Claude Fable 5.1 and Mythos 5.1 were announced. Both run the same base model with different safety layers.

• September 22, 2026: Claude Opus 5.5 was released.

• The coming weeks: Anthropic announced that Sonnet 5.5 and Haiku 5.5 will arrive "in the coming weeks" with similar improvements.

This timeline has an important consequence: three weeks after Fable 5.1 was released, a cheaper model now beats it on most work. That puts model selection back on the table for enterprises.

The price cut alone is notable, but the real impact comes from cache reads. In enterprise scenarios that work with repeated context, such as long system prompts, knowledge base content or multi-step agent flows, most of the cost comes from tokens read from the cache. A 60% drop there shows up directly on the bill. Anthropic states that on typical workloads Opus 5.5 runs about 40% cheaper than Opus 5.

One change deserves attention: in Opus 5.5, thinking mode can no longer be turned off. The model always runs with adaptive thinking; how much it thinks is set by the effort setting. The levels are low, medium (default), high, xhigh and max.

Claude Opus 5.5 benchmark results

The scores below are taken from Anthropic's launch page. The highest value in each row is in bold.

One point to keep in mind when reading the scores: a significant share of the results depends on the effort setting. For example, on Terminal-Bench 4.0 each model is listed with its own highest score; Opus 5.5 was measured at max effort, GPT-6 Astra at high effort. FrontierCode, on the other hand, was run for Opus 5.5 at the default medium setting.

Coding benchmarks

The biggest leap is in terminal-based tasks. On Terminal-Bench 4.0, Opus 5.5 leads Opus 5 by 14 points and beats Fable 5.1 by more than 10 points. The 11-point gain on CursorBench 4.0 points to a similar picture: the model completes multi-step tasks in real developer environments in fewer steps and more consistently.

Computer use and automation

On OSWorld 2.0, the gap between Opus 5.5 and Fable 5.1 is only 1.1 points. In other words, the two models are almost on par for computer use. On AutomationBench, GPT-6 Astra is still slightly ahead; however, with a 13-point gain over Opus 5, Opus 5.5 closes most of the gap.

Knowledge work and reasoning

GDPval-AA v2.1 is an Elo measurement that compares models on real work outputs. Here, Opus 5.5's 1846 points is by far the highest score. According to Anthropic, Opus 5.5 at medium effort beats GPT-6 Astra at max effort for about a fifth of the cost per task. Opus 5.5 also tops the list on Humanity's Last Exam.

Science and visual output

On Terminal-Bench-Science, Opus 5.5 scores almost twice as high as Opus 5 (from 29.0% to 58.7%), but GPT-6 Astra is still ahead in this area. On Chartography, which measures chart interpretation and generation, Opus 5.5 and Fable 5.1 are neck and neck.

What difference does it make in practice?

Benchmarks only tell part of the story. Early access feedback shows more clearly how the difference feels in day-to-day work.

Long, unattended tasks

One of the examples Anthropic shared is a 680,000-line code migration completed in less than a day. The Clio team reports that on an engineering task running unattended for 18 hours, the model stayed focused and required very little rework. Stripe says it completed a multi-day, 40-PR rebase in a single session.

Fewer steps, fewer tokens

A common observation across many companies is that the model finishes the same work in fewer steps. GitHub says the number of steps on terminal tasks dropped by more than half compared to Opus 5. Factory reports seeing 20–25% fewer output tokens on long investigations, and Rogo states that on a finance benchmark it reached Opus 5's high-effort performance with 60% fewer tokens. On the enterprise side, this means a second line of savings on top of the price cut.

A more natural writing style

One of the most common criticisms of Opus 5 was that it wrote more than necessary. This has changed noticeably with Opus 5.5. Box reports outputs that are 40% less verbose without sacrificing accuracy. Testers note that the model puts the important information at the start of the message, uses less jargon and follows writing guidelines better.

Does anything change when writing prompts?

Most prompts written for Opus 5 will keep working. Still, two points are worth noting. First, if you have flows that rely on the length and format of the output, take into account that Opus 5.5 writes more concisely. Second, since thinking mode cannot be turned off, you now need to manage cost and latency through the effort setting. The default medium setting is enough for most work; for difficult tasks, trying high or xhigh makes more sense.

Safety and alignment

Anthropic states that Opus 5.5 achieved the best score of any model tested to date on a 2,000-scenario automated behavioral audit. The model is noticeably more cautious than previous models about taking hard-to-reverse actions, and it performs the same as or better than Opus 5 against prompt injection in every scenario tested. External evaluations were carried out by METR and Frontier Design.

The line drawn on cyber and biology

Because Opus 5.5's capabilities in biology were measured at the same level as Mythos 5.1, the safeguards applied to Fable also apply to this model. A significant share of cybersecurity tasks is restricted by protection layers; verified experts can get broader access through the expanded Cyber Verification Program, and institutions such as academic labs and pharmaceutical companies through the Life Sciences Verification Program.

Two more notes on enterprise compliance: Opus 5.5 outputs use watermarking for EU AI Act compliance, and the zero data retention option remains available.

Claude family comparison: Where does each model stand?

The most striking row in this table is the relationship between Fable 5.1 and Opus 5.5. Anthropic's own model guide now also recommends starting with Opus 5.5 for most workloads, and bringing in Fable 5.1 only when Opus 5.5 falls short even at high effort. In other words, the most expensive model in the family is no longer the default choice but a fallback option.

Who is Claude Opus 5.5 for?

• Teams using Fable 5.1: In most scenarios, you can get the same or better results at a price 2.5 times lower. There is no longer a strong reason to stay on Fable without running a comparison on your own evaluation set.

• Teams working on Opus 5: A cheaper, faster model that consumes fewer tokens. Switching is the lowest-risk step.

• Teams building long-running, unattended agent flows: Terminal-Bench and the early access examples show that consistency improves noticeably on tasks that run for hours.

• Knowledge-work-heavy teams: In work such as reporting, analysis and document generation, both the GDPval score and the more natural writing style make a direct difference.

• High-volume, simple tasks: Sonnet 5 or Haiku 4.5 still make more sense here. It is also worth factoring in that Sonnet 5.5 and Haiku 5.5 are coming soon.

Conclusion

Opus 5.5 is a release that shifts the balance within Anthropic's model family. It does the work of Fable 5.1, which was positioned in the top segment only a few weeks ago, at a price even lower than Opus 5 and faster. The practical takeaway for enterprises is this: model selection can no longer be made on the assumption that "the most expensive is the best." Testing Opus 5.5 on your own workload at medium and high effort is the shortest path to improving both cost and quality at the same time.