Claude Opus 5.5 — 40% Savings, Defaults Hide the Catch

Anthropic shipped Claude Opus 5.5 on September 22, 2026 — $4 input, $20 output per million tokens, cache reads slashed to $0.20/MTok. The headline number is “40% cheaper than Opus 5 on typical workloads.” The number is real. The measurement methodology is where things get interesting: Anthropic is comparing Opus 5.5 at its medium default effort against Opus 5 at its high default effort. If you drop claude-opus-5-5 into an existing pipeline without touching the effort parameter, you are not getting the same model behavior at a lower price. You are getting a different behavior profile that happens to cost less.

TL;DR

• What: Opus 5.5 ships at $4/$20 per MTok — 20% sticker cut, 40% claimed workload savings vs Opus 5
• Catch: The 40% is measured at each model’s default effort — medium for 5.5, high for 5. Apples-to-oranges unless you explicitly set effort levels
• Breaking changes: Preserved Thinking returns 400 errors on newer accounts; thinking can no longer be disabled; forced tool use is gone
• Action: Pin effort: high if you need Opus 5-equivalent quality, audit for Preserved Thinking and safeguard rerouting before migrating

Claude Opus 5.5 — What Happened

Anthropic released Opus 5.5 as a drop-in replacement for Opus 5 across the API and Claude.ai. The pricing moves from $5/$25 to $4/$20 per million input/output tokens — a clean 20% sticker reduction. Cache reads drop 60% to $0.20/MTok, which is the real savings lever for agentic and coding workloads where cache hits dominate the bill.

The model posts 66.4% on Terminal-Bench 4.0 at extra-high effort, beating GPT-6 Astra’s 57.9% on the same benchmark. Before you frame this as a clean sweep: Anthropic’s own published numbers show Astra winning on AutomationBench (business-workflow automation) and Terminal-Bench-Science, where Astra scores 64.6% against Opus 5.5’s 58.7%. The frontier is shared, not owned.

Alongside the model swap, Anthropic changed the default effort level from high to medium. Medium on Opus 5.5 reportedly matches or beats Opus 5 at high effort on most tasks — but “most tasks” is doing a lot of heavy lifting when your pipeline has specific quality SLAs. The 40% cost reduction claim flows directly from this default shift: you are comparing a model running at reduced compute against its predecessor running at full compute.

Two breaking changes ship with the release. First, Preserved Thinking: for API accounts created on or after August 31, 2026 (00

UTC), editing prior thinking blocks out of conversation context and replaying them returns a 400 error by default. Any agent framework that rewrites prior assistant turns — and plenty do — will start failing silently or loudly depending on error handling. The mitigation is sending the thinking-binding-controls-2026-08-01 beta header and setting thinking.block_binding.prefix_mismatch_behavior to drop_block, or switching to mid-conversation system messages instead of rewriting history.

Second, adaptive thinking can no longer be disabled. Requests that explicitly set thinking: {type: "disabled"} will fail. Forced tool use also returns an error. If you were using disabled thinking as a latency or cost optimization, you now need to use a lower effort level instead.

Why This Matters

I’ve seen this movie before. Anthropic ships a cost-reduction story, buries the lever in a default setting change, and every team that migrates blind gets a surprise — either in their bill or in their output quality, depending on which direction the default moved.

The effort-level default dropping from high to medium is the actual news here, not the price cut. Here is why: most production agentic pipelines do not explicitly pin effort levels. They call the model, pass the prompt, take the output. When Opus 5 was the target, the implicit effort was high. Swapping to claude-opus-5-5 without changing anything else gives you medium-effort responses that may or may not meet the quality bar your pipeline was tuned against.

This creates two failure modes. Teams that care about cost will celebrate the 40% reduction without noticing their agent’s output quality degraded on edge cases. Teams that care about quality will pin effort: high to maintain parity — and their 40% savings evaporates back toward the 20% sticker reduction, or disappears entirely if they push to extra-high effort to chase the Terminal-Bench 4.0 numbers Anthropic is marketing.

The Preserved Thinking change is a second migration landmine. It specifically targets multi-turn agent architectures where the framework manages conversation history. If your orchestration layer strips, edits, or replays thinking blocks from prior turns — a common pattern in frameworks like LangGraph and custom agent loops — you will start seeing 400 errors on any API account created after August 31. Older accounts are grandfathered in for now, but the migration guide makes clear this is a deprecation path, not a permanent exemption.

The safeguard stack adds a third dimension to audit. Cybersecurity tasks are silently rerouted to Claude Opus 4.8. Biology-flagged requests route to Claude Opus 5. If you are running code-security agents or bioinformatics pipelines, classifier hits may be causing undetected fallback routing to older, cheaper models — without any API error or flag to surface it.

Compare this to the Fable 5.1 cache cut, which similarly bundled genuine savings with behavioral changes that required explicit migration work. The pattern is consistent: Anthropic optimizes the default path for new users and cost-conscious workloads, while existing production users absorb the migration tax.

The cache-read price drop to $0.20/MTok is the most straightforward win. For agentic workloads with high context reuse — coding assistants, document analysis pipelines, multi-step research agents — cache reads can represent 60-80% of total token volume. A 60% reduction on that line item compounds into real savings regardless of effort-level settings. This is the number to focus on if you are building a cost model.

Before migrating, run your existing test suite against Opus 5.5 at both effort: medium and effort: high. Compare output quality deltas against your actual evaluation criteria, not Anthropic’s benchmarks. The gap on your workload may be negligible — or it may be the difference between shipping and debugging.

The Take

The 40% number is marketing arithmetic. Real savings depend on three variables: your effort-level pin, your cache-hit ratio, and whether your pipeline triggers safeguard rerouting. Teams that audit all three before migrating will capture genuine cost reductions — the cache pricing alone justifies the work. Teams that swap model strings and move on will discover the gap between benchmark marketing and production reality the hard way.

My concrete recommendation: pin effort: medium explicitly (do not rely on defaults — they change), run your eval suite, and only relax to the new default behavior if your quality metrics hold. Audit your agent framework for thinking-block replay patterns. Check whether any of your workloads hit cybersecurity or biology classifiers. Then — and only then — update your cost projections.

Opus 5.5 is a better model at a better price. That part is real. But the migration path has two landmines and a hidden rerouting system, and the flagship cost claim assumes you accept a default you may not want. Read the fine print before you celebrate the savings.

Related

• Claude Opus 5 Effort Ladder Cost Trap — the original effort-level pricing analysis that predicted this pattern
• Claude Fable 5.1 Cache Cut Breaking Changes — same bundle pattern: genuine savings + behavioral changes requiring explicit migration
• Claude Code Nested Subagents Token Cost — how nested agent architectures amplify per-token pricing changes