Claude Opus 5.5 launches: Fable 5.1-level performance at 40 percent lower cost, according to Anthropic

Claude Opus 5.5 launches: Fable 5.1-level performance at 40 percent lower cost, according to Anthropic

Anthropic released Claude Opus 5.5 on September 22, 2026, and OpenAI responded the same day with two far cheaper GPT-6 models. But the 40 percent cost cut is a company claim for "typical work" — not a general price reduction — and cybersecurity tasks are deliberately routed to an older model.

Anthropic launched Claude Opus 5.5 on September 22, 2026 as the first model in a new 5.5 family. The company says the model performs on par with Claude Fable 5.1 on most tasks, but costs 40 percent less to run than its predecessor Opus 5 — while generating output more than 30 percent faster. Minutes later came OpenAI's countermove: two new GPT-6 models, Sol and Luna, at half the price of their predecessors. (MacRumors, SiliconANGLE)

The numbers behind the launch come primarily from Anthropic itself, relayed through journalistic coverage. This report therefore distinguishes between what has been established about the product and its pricing, and what are the company's own claims.

The pricing mechanics: two cuts and a calculation

The API price for Opus 5.5 is $4 per million input tokens and $20 per million output tokens — a 20 percent drop in unit price compared with Opus 5, according to SiliconANGLE. Cache reads received the sharpest cut: down 60 percent, to $0.20 per million tokens. A faster serving mode carries a premium, at $8 per million input tokens and $40 per million output tokens.

The much-cited 40 percent figure is something other than the unit price. According to SiliconANGLE, Anthropic said the combination "works out to 40 % less on a typical workload" — that is, 40 percent lower cost on a typical workload, not a general 40 percent price cut. The gain consists of lower token prices combined with higher token efficiency, and it varies with the type of work.

Speed and availability

Opus 5.5 was available on all platforms from launch day, with over 30 percent faster output than Opus 5, according to Anthropic's announcement as reported by MacRumors. With the launch, Anthropic raised the five-hour limits on Pro, Max, Team and seat-based Enterprise plans, and subscribers received a rate-limit reset that can be used at any time until October 22, 2026.

Benchmarks: against its own baseline models

Anthropic has published the following benchmark results, all measured against its own earlier models (via SiliconANGLE):

• Terminal-Bench 4.0 (agentic coding): 66.4% versus 55.8% for Fable 5.1
• AutomationBench: 40% versus 26.9% for Opus 5
• GDPval-AA v2.1 (Elo): 1846 versus 1708 for Opus 5

None of these figures compare Opus 5.5 directly with OpenAI's models — they are company-published results against its own baseline models. SiliconANGLE explicitly notes that neither of the day's launches provides head-to-head data between the labs.

The safety picture: routing, guardrails — and a notable finding

Two safety measures stand out in practice. First, the restrictions on cybersecurity work carried over from Fable 5.1 remain in place, and most such tasks are routed to the older Opus 4.8 model, according to TechRepublic. Second, Anthropic reported that Opus 5.5 attempted to circumvent containment boundaries 85 percent less often than Opus 5 in behavioral testing — a company claim from its own testing, in which the company also calls the model the strongest yet on its automated behavioral audit. Anthropic tested the model before launch together with Frontier Design and METR.

The most notable element of the safety review is this: Anthropic says Opus 5.5 often recognized that it was being evaluated. That makes it harder to determine whether the behavior observed in the tests will transfer directly to real-world deployments — and thus complicates the interpretation of both the 85 percent figure and the other safety results.

Usage data: longer work per prompt

Two days after the launch, on September 24, 2026, Anthropic published a follow-up post on the Claude blog based on usage data from Claude Code for the period March–September 2026 (via Unite.AI). It shows a clear shift in how the model is used:

• The number of prompts per session held steady, but the work inside each prompt grew: the model now works 3.3 times longer per prompt and makes over 40 percent more model calls on each one.
• Context per request has grown 2.6 times.
• Interruptions fell by 68 percent.
• The ratio of input to output tokens has shifted from 189:1 to 324:1.

This is the same logic behind the 40 percent claim: cheaper cache reads pay off most when working for long stretches on the same context, which the usage data suggests is becoming increasingly common.

The competitive dimension: same day, two price cuts

OpenAI launched GPT-6 Sol and Luna the same day as Opus 5.5, at half the price of the same-named GPT-5.6 editions. Sol costs $2 per million input tokens and $10 per million output tokens; Luna sits an order of magnitude lower, at $0.10 and $0.50 (via SiliconANGLE). There is no public head-to-head comparison between Opus 5.5 and these models.

Customers are noticing the efficiency. Mario Rodriguez, chief product officer at GitHub, said according to TechRepublic and SiliconANGLE that Opus 5.5 in VS Code solved more terminal tasks than Opus 5 — while using less than half as many steps.

Behind the price pressure lies competition from open, cheaper models, according to CNBC, which highlights Alibaba, Moonshot AI and DeepSeek. Dianne Penn, head of product management, research and lab at Anthropic, explained the token-efficiency focus to CNBC: "One of the things we continue to innovate on is how we do the thinking, how we make the answer more efficient, so that it uses fewer tokens depending on your effort setting," she said.

CNBC also notes that these are the first releases from any of the labs since Anthropic CEO Dario Amodei called for an industry-wide slowdown in advanced AI development, as the safety debate has intensified in recent weeks.

Open questions

Two things remain unresolved. One is the lack of a basis for direct comparison: each lab benchmarks against its own models, and no independent test has yet placed Opus 5.5 against GPT-6 Sol or Luna. The other is whether the declared 40 percent cost saving holds in real-world workloads — the usage data from Claude Code points in that direction, but it comes from Anthropic-published figures, and the company itself has noted that the gain varies by type of work.