Claude Opus 5.5 Launches: Fable 5.1-Level Performance at 40% Lower Cost — What's New, What Breaks, and What Thai Organizations Should Do

If you read our Claude Fable 5.1 launch coverage earlier this month, the open question then was whether you had to pay Fable prices to get Fable results. Anthropic has now answered that itself. The official documentation changed its guidance to "start with Opus 5.5 for most workloads; use Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5.5 at higher effort still fall short."

1. Where Opus 5.5 Sits in the Claude Lineup

Anthropic's documentation now lists four current models. Opus 5.5 is the new workhorse: 2.5 times cheaper than the top model with equal or better published scores.

Source: Models overview, platform.claude.com (26 Sep 2026) · Opus 5 and Opus 4.8 remain available as legacy models at the old $5 / $25 price

Note the release cadence. Opus 5 shipped on 24 July 2026, only two months before Opus 5.5, and Anthropic says Sonnet 5.5 and Haiku 5.5 "will follow in the coming weeks." If your organisation runs high-volume work on Sonnet 5, there is no reason to move it yet. Wait for its own 5.5 release.

2. Benchmarks: Beats Fable 5.1 on Every Reported Set, Trails GPT-6 Astra on Two

Anthropic published nine benchmark results alongside Fable 5.1, Opus 5 and OpenAI's models. We reproduce all of them, including the two where Opus 5.5 does not lead.

Source: Anthropic, Introducing Claude Opus 5.5 (22 Sep 2026) · all figures measured by Anthropic

Three things stand out. First, Opus 5.5 beats Fable 5.1 on all nine sets at 2.5 times lower price, which is why Anthropic is comfortable saying "Fable 5.1-level." Second, the widest gap over Opus 5 is in agentic work (Terminal-Bench 52.3 → 66.4, AutomationBench 26.9 → 40.0), not in question answering. Third, GPT-6 Astra still leads on AutomationBench and Terminal-Bench-Science. Anthropic's answer is cost: at default effort, Opus 5.5 spends about 20% of GPT-6 Astra's cost per task on FrontierCode and about 40% on Terminal-Bench 4.0.

3. Pricing: A 20% Sticker Cut That Becomes a 40% Bill Cut

Every Opus 5.5 price line is exactly 20% below Opus 5, except one that dropped much further.

Source: Pricing, platform.claude.com (26 Sep 2026) · the full 1M-token context is billed at one rate, with no long-context surcharge

So where does "40% less" come from when the sticker says 20%? Anthropic measured the cost of real workloads, not the price per token. Three things pull the workload cost down.

1. The 20% list-price cut on every line.
2. Fewer tokens per task, because output is more than 30% faster and the default effort dropped from high to medium, which means less thinking on routine work. Anthropic says the 40% figure comes from its own typical workloads. We cannot verify that part.
3. Cache reads 60% cheaper, which matters most for agents that resend the same system prompt, tools and history on every turn.

Working from list prices alone: take one agent turn that sends 1,000,000 input tokens, of which 800,000 are cache hits, and generates 100,000 output tokens.

• Opus 5: (200,000 × $5) + (800,000 × $0.50) + (100,000 × $25) per million = $1.00 + $0.40 + $2.50 = $3.90
• Opus 5.5: (200,000 × $4) + (800,000 × $0.20) + (100,000 × $20) per million = $0.80 + $0.16 + $2.00 = $2.96 (24% less)

The gap between 24% and 40% is the token reduction per task, which depends on your workload. Short questions with no caching will land near 20%. Long agent loops will land closer to Anthropic's number.

One more thing if you compare against models older than Opus 4.7: the newer tokenizer introduced with 4.7 produces roughly 30% more tokens for the same text. Any per-token price comparison across that boundary has to account for it. See our Opus 4.8 article for details.

4. On claude.ai and Claude Code: Which Plans Get It, and What Everyday Users Will Notice

For people who do not write code, the launch changes three things. Opus 5.5 appears in the model picker of every paid plan from day one. Anthropic raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans and gave subscribers a one-time rate-limit reset they can save and use whenever they choose. And the model's writing behaviour changed in ways office workers are likely to prefer.

On behaviour, Anthropic states plainly that Opus 5.5 "communicates more naturally than prior models," puts the most important information up front, uses less jargon, and follows the writing rules you give it. The developer documentation adds that it reads values off dense charts, diagrams and screenshots far more precisely without tools. If your job involves summarising reports made mostly of charts, that is the feature to try first.

If your organisation procures Claude Team or Enterprise through Grand Linux, this launch does not change your per-seat price. Plans are billed per seat per year, not per model. See the structure on our Claude pricing page and per-plan usage limits in How much can each Claude plan do?

5. For Developers: Four Things That Break Existing Opus 5 Code, and One That Goes Quiet

Anthropic lists four breaking changes for code running on Opus 5. The first three are the same ones that arrived with Fable 5.1; the fourth is new. A fifth change fails no request but silently empties a user-facing progress display.

Two further changes raise no error but will change your results. First, the default effort moved from high to medium: any request that omits effort now thinks less than it did on Opus 5. Second, at the same effort level Opus 5.5 tends to think more per turn, most of all at xhigh and max. Anthropic recommends re-running your effort sweep instead of carrying settings over, and leaving room in max_tokens for the thinking.

The minimum migration for code already running Opus 5 with thinking on is this:

# Before
response = client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    messages=[...],
)

# After: change the model ID and set effort explicitly (the new default is medium)
response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    output_config={"effort": "high"},   # pick the level your own sweep supports
    messages=[...],
)

New on the API side: fast mode for Opus 5.5 on the Claude API only, at $8 / $40 and up to 2.5 times faster; defining tools inside a mid-conversation message without losing the prompt cache (beta inline-tools-2026-09-15); compaction on demand (beta compact-2026-09-04); and per-message effort changes. Teams already on Claude Code can run /claude-api migrate to sweep a whole project instead of editing by hand.

6. Safeguards and "Pacing the Frontier": How an Organisation Should Read It

Opus 5.5 is the first model Anthropic has shipped since founder Dario Amodei wrote that "AI progress should be paced so that safety practices stay ahead of model capabilities." Many read that as a slowdown in releases. What actually happened is a new model on the usual two-month cadence, at a lower price, carrying the same risk controls as the top-tier model.

Those controls are Fable 5.1-level safeguards in two areas. For cybersecurity, most requests that trip the classifier are re-routed to Opus 4.8 automatically. For biology, deeper work requires the Life Sciences Verification Program. The model was tested by external evaluators METR and Frontier Design before release, and a full System Card is published.

From a procurement angle, what "pacing" gives an organisation is predictability. Anthropic committed not to retire Opus 5.5 before 22 September 2027, a full 12 months on the Claude API, Claude Platform on AWS and Microsoft Foundry. An annual contract signed today covers the model's guaranteed life exactly. Bedrock and Google Cloud set their own dates.

7. What Thai Organisations Should Do: A Checklist by Role

On minimum seats, a question we get often: Anthropic's Team minimum is 2 seats. Grand Linux's procurement service with Thai paperwork starts at Team 5 seats and Enterprise 20 seats, up to 150. Teams of 2 to 4 are better off buying directly from Anthropic, and we wrote up how.

8. The ERP View: Models Change Every Two Months, the Source of Truth Must Not

Opus 5 held the "latest" label for two months. Opus 4.8 before it did not last much longer. The lesson for any organisation running an ERP (Enterprise Resource Planning) system is to bind your processes to an integration layer that swaps models with a single setting, never to one model.

Saeree ERP ships with no built-in AI feature, and that is deliberate. The ERP is the source of truth, running on the customer's own infrastructure. What Grand Linux builds is the connection between Claude and ERP data over MCP (Model Context Protocol), with permissions scoped to the user's role and every call logged. When Anthropic ships a new model, moving from Opus 5 to Opus 5.5 is a one-line model ID change in that layer. The breaking-change table in section 5 is exactly what our team checks on our own integrations before recommending anything to customers. Read the concept in What is MCP? and the integration patterns on Claude Solutions.

Data deserves its own sentence. ERP data lives on the organisation's servers or on cloud it chose; Claude runs on Anthropic's cloud or a chosen cloud provider. An MCP integration therefore has to define which data may be sent out to be asked about and which stays inside. A smarter model does not change that principle.

Conclusion

• Claude Opus 5.5 launched on 22 September 2026 at $4 / $20 per million tokens, 20% below Opus 5, with cache reads 60% cheaper; Anthropic measures real-workload cost 40% lower.
• It beats Fable 5.1 on all nine reported benchmarks at 2.5 times lower price, and the documentation now says "start with Opus 5.5."
• Pro, Max, Team and Enterprise get it immediately, with higher five-hour limits and no change to seat prices.
• Developers must check four breaking changes, the new default effort of medium, and thinking.display for progress displays.
• Sonnet 5.5 and Haiku 5.5 follow in weeks; high-volume Sonnet 5 workloads can wait.
• Bind work to the integration layer, not the model. The source of truth stays in the ERP.

The best model for an organisation is not the smartest one in the table. It is the one that finishes your work at a cost you can explain to accounting, and that you can swap without rebuilding the system.

- Saeree ERP team

References

• Anthropic — Introducing Claude Opus 5.5 (22 Sep 2026)
• Claude Platform Docs — Claude Opus 5.5 model page (specs, pricing, platforms)
• Claude Platform Docs — What's new in Claude Opus 5.5 (breaking changes and behaviour)
• Claude Platform Docs — Migrating to Claude Opus 5.5
• Claude Platform Docs — Pricing (checked 26 Sep 2026)
• Claude Platform Docs — Models overview
• TechCrunch — Anthropic releases Opus 5.5 with lower prices and Fable-level performance (22 Sep 2026)