cd /news/large-language-models/claude-opus-5-5-40-cheaper-30-faster… · home topics large-language-models article
[ARTICLE · art-137380] src=byteiota.com ↗ pub= topic=large-language-models verified=true sentiment=↑ positive

Claude Opus 5.5: 40% Cheaper, 30% Faster, Beats Fable

Anthropic released Claude Opus 5.5, cutting typical workload costs 40% versus Opus 5 with input tokens at $4 per million (down from $5), output at $20 (down from $25), and cache reads at $0.20 (down from $0.50, a 60% reduction). Anthropic said Opus 5.5 outperforms its larger premium model Fable 5.1 on most benchmarks, including Terminal-Bench 4.0 (66.4% vs 55.8%), FrontierCode v1.1 (54.4% vs 50.3%), and AutomationBench (40.0% vs 31.4%), while generating output 30% faster and completing tasks in 40-50% fewer turns. Opus 5.5 ships with safety classifiers that silently reroute cybersecurity requests to Opus 4.8 and require Life Sciences Verification Program enrollment for biology tasks, a concern for multi-turn agent workflows where steps may run on different model versions.

read4 min views5 publishedSep 22, 2026
Claude Opus 5.5: 40% Cheaper, 30% Faster, Beats Fable
Image: Byteiota (auto-discovered)

Anthropic shipped Claude Opus 5.5 today, and the headline number is 40% — that’s how much less you’ll spend per typical workload compared to Opus 5. Input tokens drop from $5 to $4 per million, output from $25 to $20, and cache reads fall from $0.50 to $0.20. That last one is the number to watch if you’re running agents.

The Cache Math Agents Actually Care About #

On a standard API request, a 20% price cut is fine. On an agentic coding workflow, the savings stack differently. Agents burn cache reads aggressively — every turn in a multi-step task reuses the same large system prompt and accumulated context. Cache reads used to cost $0.50 per million tokens. They now cost $0.20. That’s a 60% reduction on the line item that often dominates agent billing.

Combined with Opus 5.5 completing tasks in 40–50% fewer turns and generating output 30% faster than Opus 5, the practical cost reduction on real agent workloads is significant. Optiver confirmed this in Anthropic’s release materials: their agentic coding workflows saw 40–50% cost reduction at the same quality, using roughly half the turns. Box AI reported 40% fewer output tokens with no accuracy loss. One Clio developer ran a multi-repository task for over 18 hours without the model losing track of the goal.

| Token Type | Opus 5.5 | Opus 5 | Change |

|---|---|---|---|
| Input (per 1M) | $4.00 | $5.00 | -20% | 
| Output (per 1M) | $20.00 | $25.00 | -20% | 
| Cache reads (per 1M) | $0.20 | $0.50 | -60% | 
| Cache writes (per 1M) | $5.00 | $6.25 | -20% | 

| Typical workload | — | | -40% |

It Beats the Larger Model #

Anthropic’s benchmark headline is that Opus 5.5 outperforms Fable 5.1 — the company’s larger, more expensive premium model — across most evaluations. On Terminal-Bench 4.0, Opus 5.5 scores 66.4% against Fable’s 55.8%. On FrontierCode v1.1, 54.4% against 50.3%. On AutomationBench, 40.0% against 31.4%. The knowledge work Elo gap (1,846 vs 1,735) also shows up in finance tasks — Hebbia reported 86.6% coverage of expert-graded criteria versus Opus 5’s 60.3%.

Beating a larger model at a lower price tier is the real story here. Anthropic’s Fable lineup sits at the top of their pricing. Opus 5.5 performing comparably changes how teams should structure their model routing — there are now fewer reasons to reach for Fable unless you specifically need something it still leads on.

One Catch: Your Request Might Not Land on Opus 5.5 #

This is the part most migration guides will skip. Opus 5.5 ships with safety classifiers that silently reroute certain requests to older models when they fire. Most cybersecurity tasks — exploit analysis, vulnerability research, penetration testing — get rerouted to Opus 4.8, not Opus 5.5. Biology tasks require enrollment in the Life Sciences Verification Program.

In a single API call, that’s manageable. In a multi-turn agent workflow, it creates an architectural concern: individual steps may be handled by different model versions with different capabilities. A step processed by Opus 4.8 behaves differently than one processed by Opus 5.5, and that inconsistency affects downstream steps.

If your agent does anything that could trigger a security classifier — code auditing, dependency scanning, threat modeling — test with response metadata logging to confirm which model is actually handling each request. [The New Stack flagged this concern](https://thenewstack.io/claude-opus-5-5-release/) at launch. Anthropic is expanding its Cyber Verification Program to Opus 5.5 for vetted security teams. Apply before you’ve already built the workflow around it.

## Safety Context (Brief)

Opus 5.5 is Anthropic’s first model release since CEO Dario Amodei publicly committed to pacing the frontier — deliberately slowing capability advancement to give alignment research time to catch up. The model was evaluated by METR and Frontier Design before shipping. On containment evaluations, Opus 5.5 recorded 85% fewer boundary violation attempts than Opus 5. EU AI Act watermarking is included.

What to Do Now #

The model ID is claude-opus-5-5. It’s live on the Claude Platform, AWS, Google Cloud, and Azure. Swapping the model ID from claude-opus-5 gets you the pricing improvements immediately. Test agent workflows before fully migrating — particularly any involving security-adjacent tasks. TechCrunch covered the full release including benchmark details.

Sonnet 5.5 and Haiku 5.5 are arriving within weeks. Developers who work primarily at the Sonnet price tier can expect the same generation of improvements.

The cache read price cut is real. If agents are part of your stack, calculate what the 60% cache read reduction means for your specific workload before assuming the 40% headline figure applies — pure inference jobs with no caching will see closer to 20% savings. Context-heavy agent tasks will see considerably more. Artificial Analysis has independent benchmark comparisons if you want a second opinion on the performance claims.

── more in #large-language-models 4 stories · sorted by recency
── more on @anthropic 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/claude-opus-5-5-40-c…] indexed:0 read:4min 2026-09-22 ·