Claude Sonnet 5.5 Launches: 30% Faster at the Same Price, Up to 30% Lower Cost per Task — What's Better, What Breaks, and How It Compares to Opus 5.5

Six days after Claude Opus 5.5 launched, Anthropic shipped Claude Sonnet 5.5 on Monday, 28 September 2026, well ahead of the "coming weeks" it had promised. For most Thai organizations there is really one question, because high-volume work runs on Sonnet 5, not Opus: if the price is unchanged, should you move, and what do you get? This article answers that step by step, then closes with a side-by-side against the Opus 5.5 that arrived only days earlier.

1. Where Sonnet 5.5 Sits in the Claude Lineup

Anthropic's "current lineup" page now lists four models. Sonnet 5.5 takes Sonnet 5's slot as the mid-tier model Anthropic calls "the best combination of speed and intelligence." Every price is unchanged, the knowledge cutoff moves forward five months, and the scores land far closer to Opus 5.5 than Sonnet 5 ever got to Opus 5.

Source: Models overview, platform.claude.com (29 Sep 2026) · Sonnet 5 remains available as a legacy model at the same $2 / $10 with a January 2026 knowledge cutoff

Two rows matter more than any score. First, the knowledge cutoff moves from January to June 2026, the same as Opus 5.5 and Fable 5.1. Questions about regulations, prices, or events from the first half of this year get answered without a web search. Second, the tokenizer is the same as Sonnet 5's. The same text produces the same token count. When Sonnet 5 launched, every cost model had to be redone because its new tokenizer counted roughly 30% more tokens. This time there is nothing to redo.

Haiku 5.5 is not here yet. Anthropic says it "will join the Claude 5.5 family in the coming weeks." Haiku 4.5 retires on 15 October 2026, so systems still on it have under three weeks. If Haiku 5.5 does not arrive in time, Sonnet 5.5 at low effort is the option Anthropic recommends for short, high-volume replies.

2. What Improved over Sonnet 5: Anthropic's Numbers and Customers' Numbers

Anthropic's own summary is short: "a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work." The second 30% does not come from the price list, which did not change. It comes from using fewer tokens and fewer tool calls to finish the same piece of work. The table below groups the improvements with the figures that can be checked.

Source: Introducing Claude Sonnet 5.5 (28 Sep 2026) and What's new in Claude Sonnet 5.5 · customer figures are each company's own statement in the announcement

The customers Anthropic quotes all say the same thing: speed on routine work. Zendesk reports tickets processed 20% faster than its current production model. Box says data-accuracy tasks run 2.4 times faster than on Sonnet 5. Atlassian says its Rovo Agents run up to 30% faster. These are each company's numbers on its own workload. Treat them as a guide to what to test first, not as results you will reproduce.

3. Where It Is Strongest: Reading All Eight Benchmarks

Anthropic published eight benchmarks comparing Sonnet 5.5 with Sonnet 5, Opus 5.5, and OpenAI's models. We include all of them, including the ones where Sonnet 5.5 still trails.

Source: Introducing Claude Sonnet 5.5 (28 Sep 2026) · all figures are Anthropic's own · Anthropic notes that Sonnet 5.5's GDPval-AA and AA-Briefcase scores come from a pre-release build with a since-fixed structured-outputs bug, so the true figures are likely slightly higher

Three things stand out. First, on six of the eight suites Sonnet 5.5 trails Opus 5.5 by three points or less at half the price. On office knowledge work the gap is 2 Elo points, which is no gap at all. Second, the gaps that remain are on hard coding problems (FrontierCode, 8 points) and deep academic knowledge (Humanity's Last Exam, 3 points). That is exactly the work Anthropic still steers to Opus. Third, the Terminal-Bench 4.0 result where Sonnet 5.5 beats Opus 5.5 needs its footnote: the Opus 5.5 figure in the table is its xhigh-effort result, not max, so it is not a like-for-like comparison.

The first independent measurement, from Artificial Analysis on launch day, tells the same story. Its overall Intelligence Index puts Sonnet 5.5 at 56 and Opus 5.5 at 58. By domain, Opus leads in finance and accounting (61 to 57) and engineering (60 to 58). Cost per task at low effort came out at $0.41 for Sonnet 5.5 against $0.55 for Opus 5.5, a 25% difference despite a 2x difference in price per token, because the larger model tends to use fewer tokens for the same job. Those are Artificial Analysis's measurements on its own test set, not Anthropic's.

4. Pricing: Every Line Unchanged, So Where Does "30% Cheaper" Come From?

Anthropic's What's new page confirms that Sonnet 5.5 has the same prices as Sonnet 5 on every line, including caching and batch rates. The table sets it against Opus 5.5 so you can see exactly where the difference between the two models sits.

Source: Pricing, platform.claude.com (29 Sep 2026) · the full 1M context is billed at one rate · the Sonnet 5 increase to $3 / $15 once scheduled for 1 Sep 2026 has been cancelled

The most interesting cell is cache reads, priced identically on both models at $0.20. Opus 5.5 cut its cache-read rate to 5% of input while Sonnet still charges 10%. For agent workloads that resend the same system prompt, tools, and history every turn, the gap between the two models is therefore less than 2x. Using the same example as our Opus 5.5 article, take one agent turn that sends 1,000,000 input tokens, 800,000 of them read from cache, and generates 100,000 output tokens.

• Sonnet 5.5: (200,000 × $2) + (800,000 × $0.20) + (100,000 × $10) per million = $0.40 + $0.16 + $1.00 = $1.56
• Opus 5.5: (200,000 × $4) + (800,000 × $0.20) + (100,000 × $20) per million = $0.80 + $0.16 + $2.00 = $2.96 (1.9x, not 2x)

The "up to 30% lower cost per task" versus Sonnet 5 is not on the price list at all. It comes from fewer tokens per task and fewer tool calls. Anthropic measured it on the typical workloads it tests internally, and we cannot verify that part ourselves. But Balyasny's figure of 121k tokens per answer, down from 497k, on its own finance tasks suggests that for work that reads long documents and summarizes them, the saving can exceed 30%. If your workload is short questions with no tools, expect the saving to approach zero, because there is nothing to economize.

5. claude.ai and Claude Code: Who Gets It, and a Default That Differs from the API

For everyday users, the announcement says Sonnet 5.5 is available on claude.ai from launch day, and Claude Code separately announced it was live the same day. Unlike the Opus 5.5 launch, though, this announcement says nothing about raised usage limits, and it does not say whether Sonnet 5.5 replaces Sonnet 5 as the default model on the Free and Pro plans.

The different default effort between the web app and the API is new and worth knowing. Anthropic set claude.ai to medium so replies come back quickly, and the API to high so quality is maximized without configuration. If users complain that the web app "thinks less" than what developers see through the API, that is the reason. It is the same model.

If your organization procures Claude Team or Enterprise through Grand Linux, the per-seat price does not change with this news. The price structure is on our Claude pricing page, and per-plan usage limits are covered in How much can each Claude plan actually use? For a broader guide to choosing between plans, see Team vs Enterprise.

6. For Development Teams: Five Things That Break, and One That Goes Quiet

Anthropic lists five breaking changes for code running on Sonnet 5. The first four are the same set that arrived with Opus 5.5. The fifth is specific to Sonnet. One more change fails no request but leaves the user's screen silent.

Source: What's new in Claude Sonnet 5.5 and the Migration guide, platform.claude.com (29 Sep 2026)

Two more changes are not errors but affect results. First, effort levels are recalibrated. The same level does not produce the same amount of thinking, and Anthropic tells you to re-run your effort sweep rather than carry the old setting over. Recommended starting points are high for general work, medium for agentic work with a clear spec, and medium or low for latency-sensitive chat. At low, the model may skip verifying its own work and is more likely to stop mid-task to check in. Second, thinking tokens count toward max_tokens even when the thinking content is not returned. A tight max_tokens left over from the thinking-off days will truncate replies. For agentic coding Anthropic recommends 128,000 with streaming.

The minimal migration for code already on Sonnet 5 with thinking on is a one-line model swap. Code that turned thinking off has to move to between_tools:

# Before: thinking off for speed
response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=4096,
    thinking={"type": "disabled"},
    messages=[...],
)

# After: new model ID, between_tools instead of disabled, room in max_tokens
response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=16000,
    thinking={"type": "between_tools"},   # accepted at high effort or below only
    output_config={"effort": "medium"},   # pick from your own effort sweep
    messages=[...],
)

API features new to Sonnet 5.5 that Sonnet 5 never had include per-message effort changes that keep the prompt cache (beta), mid-conversation system messages, tool definitions inside a mid-conversation message (beta inline-tools-2026-09-15), and compaction on demand (beta compact-2026-09-04). Teams already using Claude Code can run /claude-api migrate this project to claude-sonnet-5-5 to have the whole project checked instead of editing by hand. And because the tokenizer matches Sonnet 5, there is no token re-baselining to do.

7. Safeguards: The First Sonnet with the Same Protections as the Top Models

What international coverage focused on most was not the scores. Sonnet 5.5 is the first mid-tier model on which Anthropic has applied the same frontier-level cybersecurity safeguards as Opus 5.5 and Fable 5.1. Anthropic's reason is that the model's cyber capabilities are now "comparable to those of Opus 5," so it needs the same protection.

Two more mechanisms ship alongside. First, classifiers that block attempts to extract the model's internal reasoning, a response to the August incident in which researchers decoded hundreds of thousands of thinking blocks from publicly posted agent traces. Prompts that instruct the model to "show your reasoning in the answer" now risk a refusal; read reasoning through display: "summarized" instead. Second, Sonnet 5.5 thinking blocks work only in the account that produced them. Blocks sent from another account are dropped before they reach the model.

On behavior, Anthropic reports that across roughly 1,850 audit scenarios Sonnet 5.5 matched or improved on Sonnet 5 on most measures, and that it is "the least likely of any of our models" to probe the limits of the container it runs in. From a procurement angle, the minimum retirement date of 28 September 2027 guarantees a full 12 months on the Claude API, Claude Platform on AWS, and Microsoft Foundry. An annual contract signed today covers the model's committed lifetime exactly.

8. Versus Opus 5.5, Released Six Days Earlier: Which One to Pick

Anthropic's own framing is that Sonnet 5.5 is "a faster, lower-cost complement to Claude Opus 5.5," where Opus 5.5 is "built for complex work requiring careful judgment" and Sonnet 5.5 is "strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets." The table puts every verifiable difference in one place.

If there is one rule: start on Sonnet 5.5 at high effort and move to Opus 5.5 only for work that fails your tests there. From the table, that is three kinds of work: hard coding problems that need several layers of reasoning, open-ended tasks where the model makes dozens of its own decisions, and work that depends on deep domain knowledge. For everything else, which is most office work, the numbers say paying 2x for 2 Elo points is not worth it. A third pattern Anthropic supports directly is to run Sonnet 5.5 as the worker and call Opus 5.5 as an advisor through the advisor tool only when it gets stuck.

On the seat-minimum question we get asked most: Anthropic sets the Team minimum at 2 seats. Grand Linux's procurement service, with Thai-language paperwork, starts at Team 5 seats and Enterprise 20 seats, up to 150. Teams of 2 to 4 people are better off buying directly from Anthropic. Teams of five or more can start with our Team / Enterprise FAQ.

9. The ERP View: The Mid-Tier Model Is the One Organizations Actually Run

In the integration work Grand Linux does for customers, the model that runs day-to-day work is the Sonnet tier, not Opus. Most questions users ask an ERP (Enterprise Resource Planning) system are well-scoped: the status of a purchase order, budget versus actual spend, or a report that is nothing but charts. This launch therefore affects real costs more than last week's Opus 5.5 news did.

Saeree ERP has no AI feature built in, and that is deliberate. The ERP is the system of record, on the customer's own servers. What Grand Linux builds is the connection between Claude and ERP data over MCP (Model Context Protocol), with permissions by user role and an audit log of every call. Moving that layer from Sonnet 5 to Sonnet 5.5 is a one-line model ID change plus a pass through the breaking-change table in section 6, which our team runs itself before recommending it to any customer. Read about the approach in What is MCP? and the integration patterns on Claude Solutions.

Data has to be stated separately. ERP data stays on the organization's servers or the cloud it chose. Claude runs on Anthropic's cloud or the cloud provider you select. An MCP connection therefore has to be designed around which data may leave to be queried and which must stay in place. A faster model changes none of that.

Conclusion

• Claude Sonnet 5.5 launched on 28 September 2026 at Sonnet 5's unchanged $2 / $10 per million tokens, generates output more than 30% faster, and costs up to 30% less per task through fewer tokens and fewer tool calls.
• The clearest gains are in documents, slides, spreadsheets, chart reading, computer use from screenshots, and agentic work. Knowledge now runs to June 2026.
• On six of eight benchmarks it trails Opus 5.5 by three points or less. The remaining gaps are hard coding and deep knowledge.
• Available immediately on claude.ai, Claude Code, the API, and all three clouds. Team / Enterprise seat prices are unchanged. Default effort is medium on the web and high on the API.
• Developers must check five breaking changes, above all replacing disabled with between_tools, and re-run their effort sweep. The tokenizer is unchanged, so no token re-baselining.
• Rule of thumb: start on Sonnet 5.5 at high effort; move to Opus 5.5 only where tests fail. Haiku 5.5 follows in the coming weeks.

A mid-tier model that matches the top tier on office work at half the price moves an organization's real bill more than any new flagship does.

- Saeree ERP team

References

• Anthropic — Introducing Claude Sonnet 5.5 (28 Sep 2026)
• Claude Platform Docs — What's new in Claude Sonnet 5.5
• Claude Platform Docs — Migrating to Claude Sonnet 5.5
• Claude Platform Docs — Prompting Claude Sonnet 5.5
• Claude Platform Docs — Models overview
• Claude Platform Docs — Pricing
• Artificial Analysis — Claude Sonnet 5.5 vs Claude Opus 5.5
• TechCrunch — Anthropic releases Sonnet 5.5, which it calls a significantly cheaper, faster work partner (28 Sep 2026)
• The Decoder — Anthropic's Claude Sonnet 5.5 nearly matches Opus 5.5 on benchmarks while costing up to 30 percent less per task
• SiliconANGLE — Anthropic debuts Claude Sonnet 5.5 running 30% faster than the previous generation (28 Sep 2026)
• TNW — Anthropic releases Claude Sonnet 5.5 with the cyber limits it reserved for flagship models