Claude Opus 5.5: Flagship-Level Performance at a Lower Price
Anthropic released Claude Opus 5.5 on September 22, 2026 — the same day OpenAI launched GPT-6 Sol and Luna. It performs at the level of Anthropic's larger Fable model on most coding and knowledge work, but costs about 40% less to run than its predecessor, Opus 5. Here is what changed and why it matters.
What actually launched
Anthropic's Claude lineup has three tiers: Opus (most capable, most expensive), Sonnet (middle), and Haiku (fastest and cheapest). Opus 5.5 is the new top of that ladder.
The headline claims from Anthropic:
- It sets a new company record in coding and knowledge-work performance.
- It beats the larger Claude Fable 5.1 in many benchmarks — and even completed some informal tasks that Fable failed.
- Sonnet 5.5 and Haiku 5.5 will follow "in the coming weeks" with similar improvements.
This is also Anthropic's first model release since CEO Dario Amodei publicly called for "pacing the frontier" — deliberately slowing capability growth so safety work can keep up.
The numbers, in plain terms
- Input price (per 1M tokens) · Opus 5 (old): $5.00 · Opus 5.5 (new): $4.00
- Output price (per 1M tokens) · Opus 5 (old): $25.00 · Opus 5.5 (new): $20.00
- Typical running cost · Opus 5 (old): baseline · Opus 5.5 (new): ~40% less
Two things are happening at once, and both save money:
- Each token costs 20% less ($4/$20 vs. $5/$25).
- The model needs fewer tokens to finish the same job, which pushes the real-world saving to about 40%.
Think of it like a car that is both cheaper per gallon and goes further per gallon. Anthropic also says Opus 5.5 runs faster, uses less jargon, and puts the important information at the start of its answers — small quality-of-life upgrades that regular users will actually feel.
Where you can use it: the Claude platform, Amazon Bedrock (AWS), Google Cloud's Vertex AI, and Microsoft Foundry (Azure).
Why "cost per task" is the real story
The AI industry is quietly changing what it competes on. It used to be all about benchmark scores — who is smartest. Now the fight is about how much useful work you get per dollar.
Opus 5.5 is the clearest example yet: flagship-level results at a mid-tier bill. For businesses running AI agents that take hundreds of steps — customer support, research, coding assistants — a 40% cost drop can be the difference between "interesting demo" and "profitable product."
A note on safety
Anthropic says Opus 5.5 is comparable to its Mythos model in biology and cybersecurity capabilities, so it ships with the same safeguards as the Fable line — limits on things like discovering software exploits or helping with bioweapons. Outside evaluators including METR reviewed it before release.
What to be careful about
- "Fable-level" is Anthropic's own testing. Independent evaluations will give the fuller picture — always wait for those before treating a vendor's chart as fact.
- The price war is real, and confusing. Opus 5.5 ($4/$20) now sits above GPT-6 Sol ($2/$10) on raw token price — but below it on claimed capability. Comparing models by price alone is like comparing cars by fuel price: it misses the point.
- For most beginners, this changes little day-to-day. You will feel it indirectly, as the apps and assistants you use get better without getting more expensive.
Explain it to a friend: "Anthropic's smartest AI got a big discount. The new Claude Opus 5.5 is as capable as their huge Fable model on most work, but costs 40% less to run than the previous version. Same brains, much smaller bill."
Try this: The next time you read an AI announcement, ask one question first: did an independent tester confirm this, or only the company selling it? That single habit filters out most hype.