Claude Opus 5.5 Release Guide: Fable 5.1-Level Coding at $4/$20, and How to Run It in OpenCode

Anthropic introduced Claude Opus 5.5 on September 22, 2026, the first model in a new Claude 5.5 family that will also include Sonnet 5.5 and Haiku 5.5 in the coming weeks. The pitch in two numbers: it performs at the level of Claude Fable 5.1 on most work, and it costs 40% less to run than Opus 5. The API price is $4 per million input tokens and $20 per million output tokens (down 20% from Opus 5's $5/$25), cache reads drop 60% to $0.20 per million, and output is more than 30% faster. It is Anthropic's first release since the call to pace the frontier, with pre-release evaluation by METR and Frontier Design.

This post covers what shipped, what the benchmark table actually says, the pricing math for agent workloads, and how to run it in OpenCode right now.

Official Sources#

What shipped#

The announcement is one long efficiency story. At default effort on FrontierCode v1.1, Opus 5.5 scores 54.6%, beating GPT-6 Astra's top score (53.3%) at roughly 20% of the cost per task; on Terminal-Bench 4.0 it matches Astra for about 40% of the cost; on CursorBench 4.0 it beats GPT-5.6 Sol's best score (41.7%) by 11 points for about a third of the cost.

The other theme is writing. Anthropic says its "most common feedback" about Opus 5 was verbosity, and 5.5 is its answer: clearer messages, important information up front, less jargon. Its side-by-side samples show the same bug explanation and the same Slack-thread summary rendered in roughly half the words.

Vendor-reported early-tester results frame the ceiling: a 680,000-line code migration in less than a day, a 200,000-line codebase audited and fixed in under three hours, and a HAProxy C-to-Rust translation in 9.5 hours versus 12 for Fable 5.1 at 51% lower cost. Treat those as testimonials, not evals - the reproducible signals are the benchmark table and price list below.

Running it in OpenCode#

Opus 5.5 is available in OpenCode today as opencode/claude-opus-5-5. Install OpenCode with the official one-liner:

curl -fsSL https://opencode.ai/install | bash

Then run it, choosing an effort variant from low, medium, high, xhigh, and max:

# Long-horizon agent task at max effort
opencode run --model opencode/claude-opus-5-5 --variant max \
  "find the flakiest test in this repo and explain why it fails"

# Default-effort session for a review pass
opencode --model opencode/claude-opus-5-5

The variants map to the same adaptive-thinking settings the benchmarks use, so medium is the default-effort anchor of the announcement's cost charts. The efficiency shape is the same one we broke down for the DeepSeek V4 Flash setup our own automations run: low or medium for quick edits and review passes, xhigh or max for long multi-step loops, and cache reads at $0.20 for repo-wide context.

Decision guide#

Move to Opus 5.5 when:

• You run Claude Code or agent loops against Opus 5 today. Same family, better communication, 20% lower prices, 60% cheaper cache reads, faster output. No structural reason to stay on Opus 5.
• You need Fable 5.1-grade results without Fable 5.1 pricing. If Fable was justified only by its ceiling, benchmark the gap on your own tasks: Anthropic's framing is that the gap is narrower than scores suggest.
• Long autonomous runs are your workload: the containment-boundary improvements (about 85% fewer crossing attempts than Opus 5 or Mythos 5.1, vendor-reported) plus the action classifier matter for unattended sessions.

Stay on Fable 5.1 or wait when:

• Your hardest tasks genuinely hit Fable's ceiling. Opus 5.5 is not a Fable replacement there.
• Your workload is light: Haiku 5.5 is coming, and the calculus changes once it lands.
• Safety-verified workflows matter: under the safeguards, most cybersecurity tasks are rerouted to Opus 4.8 and biology-adjacent work to Opus 5, with verification programs for vetted organizations.

What developers are saying#

The dominant thread is relief about writing style, delayed by skepticism. Opus 5's verbosity was a live complaint through the summer - teams describing real sessions where output was hard to follow - and the community read the communication section of the announcement as directly answering that feedback. Early hands-on impressions reported drier output, though a fair share of comments said one sample does not prove a systematic change.

The second theme is pricing and positioning confusion. Why does Fable still make sense at $10/$50 when the cheaper model claims to perform at its level on most work? The counterpoint: Fable's ceiling is the differentiation, and the price only makes sense for workloads that hit it.

The third theme is skepticism about the claim cycle itself. Commenters recalled that Opus 5 was also announced as Fable-adjacent, benchmark fatigue is real, and several said they will believe the cost-per-task charts only after running the model on their own repos.

My take#

This release is the clearest signal yet that the efficiency race outran the capability race at the frontier. When a vendor leads its launch with cost-per-task charts instead of benchmark deltas, and its pricing page shows a 2.5x gap between a model and its own flagship that it claims to match on most work, the unit that matters has shifted: cost per completed task, not benchmark score.

The 5.5 family structure matters too. Sonnet 5.5 and Haiku 5.5 follow "in coming weeks" with the same efficiency improvements, so Anthropic is applying the compression play across the whole lineup - the same playbook OpenAI is running with its Luna and Sol price cuts. Benchmark leadership alone no longer converts; cheaper, faster, clearer does.

The safety posture cuts both ways. Preserved thinking blocks a class of prompt-injection extraction, and the containment-boundary improvements are aimed directly at unattended agent runs, which improves parallel-agent economics. But safeguard rerouting means capabilities are not uniformly exposed: cyber tasks fall back to Opus 4.8 and biology to Opus 5 on the standard API. Read the system card before assuming you get the full model.

Bottom line: on Opus 5 or pricing-constrained on Fable, this is the upgrade; on a cheaper tier, wait for Sonnet 5.5 and Haiku 5.5. Run it on your own repos and compare cost per completed task.

FAQ#

What is Claude Opus 5.5?#

Anthropic's new flagship-tier model, released September 22, 2026 as the first model in the Claude 5.5 family. It targets Fable 5.1-level performance on most work at 40% lower running cost than Opus 5, priced at $4/$20 per million tokens.

How do I run Claude Opus 5.5 in OpenCode?#

Install OpenCode (curl -fsSL https://opencode.ai/install | bash), then run opencode run --model opencode/claude-opus-5-5 --variant max "your task". Effort variants are low, medium, high, xhigh, and max.

Is thinking mode required?#

Yes. Thinking cannot be disabled on Opus 5.5, per the platform what's-new documentation, and preserved thinking blocks editing prior context in API sessions.

Sources#

Last updated: September 23, 2026

Continue Reading#

• Claude Opus 5 vs Opus 4.8 vs Fable 5: Full Benchmark Comparison - the July 2026 shootout this release resets
• Claude Opus 5 in 8 Minutes - the previous generation's launch guide, for the before/after
• What Parallel Claude Agents Actually Cost - the cost-per-task math behind agent fleets
• GPT-5.6 Luna 80% Price Cut - the OpenAI price move this release answers
• Frontier Model API Pricing (July 2026) - where Opus 5.5 sits against the whole field