TL;DR: Claude Sonnet 5.5 (released September 28, 2026) replaces Sonnet 5 at the identical $2/$10 per million tokens — and on Anthropic’s own launch table it beats the $4/$20 Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%) while trailing it by 2.3 points on CursorBench and 8.2 on FrontierCode. Anthropic claims 30%+ faster output and up to 30% lower cost per task from fewer tokens, not lower rates. Cursor has it in the model picker; Claude Code v2.1.284 makes it the default Sonnet on API billing; GitHub Copilot shipped it to all paid plans the same day. Cline’s built-in Anthropic catalog was still pre-5.5 at our last source check — the OpenRouter route works today.

	Claude Sonnet 5.5	Claude Opus 5.5	Claude Sonnet 5	Claude Fable 5.1
Input / output per MTok	$2 / $10	$4 / $20	$2 / $10	$10 / $50
Cache read per MTok	$0.20	$0.20	$0.20	$0.25
Terminal-Bench 4.0 (vendor-run)	70.6%	66.4%	10.3%	55.8%
FrontierCode 1.1 (vendor-run)	46.2% (at max)	54.4%	42.4%	50.3%
Comparative latency	Fast	Moderate	Fast	Slower
The catch	5 breaking API changes; loses the hardest-problem benchmark	2x the price for a mixed benchmark edge	Superseded at the same price — no reason left	Loses to both 5.5 models on coding at 2.5-5x the price

Honest take: If your agentic coding bill runs through an Anthropic API key, Sonnet 5.5 is the new default and Opus 5.5 is the model you escalate to, not the one you start with. Same price as Sonnet 5, roughly Opus-class agentic scores, and the token-efficiency gains land directly on your invoice. Keep Opus 5.5 for the hardest 20% — FrontierCode says the gap there is real — and stop paying for Fable 5.1 coding sessions entirely.

What is Claude Sonnet 5.5 and what changed from Sonnet 5?

Claude Sonnet 5.5 is Anthropic’s new mid-tier model, released September 28, 2026, six days after Opus 5.5. The API model ID is claude-sonnet-5-5 (anthropic.claude-sonnet-5-5 on Amazon Bedrock). Specs verified against Anthropic’s model docs on September 29, 2026: 1M-token context window, 128K max output (300K on the Batch API with the output-300k-2026-03-24 beta header), knowledge cutoff June 2026, retirement no sooner than September 28, 2027. It is live on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.

The pitch is different from most point releases. Per-token pricing did not move at all — the change is that the model finishes the same work with far fewer tokens:

1. Output is 30%+ faster than Sonnet 5, per Anthropic’s launch material. Comparative latency stays “Fast,” a tier above Opus 5.5’s “Moderate.”
2. Up to 30% cheaper per task, at identical rates. The savings come from token count. Anthropic’s launch page cites Balyasny Asset Management measuring about 121K tokens per answer where Sonnet 5 used 497K on the same finance tasks, and Box reporting 12% fewer total tokens. Your mileage is workload-shaped; the mechanism is real.
3. The agentic benchmark jump is enormous. Terminal-Bench 4.0 goes from 10.3% (Sonnet 5) to 70.6%. CursorBench 4.0 goes from 34.1% to 55.5%. More on why the Sonnet 5 numbers look so low below.
4. Five breaking API changes for BYOK harnesses coming from Sonnet 5 — covered in the migration section, and one of them (temperature rejection) will bite more Cline configs than you’d expect.

A pricing footnote that explains a number you may see elsewhere: Sonnet 5’s $2/$10 launched as introductory pricing with a scheduled increase to $3/$15 on September 1, 2026. Anthropic’s pricing page now states that increase never happened — $2/$10 is the standard price for both Sonnet 5 and 5.5. Any article quoting $3/$15 for a Sonnet-class model is citing a plan that was cancelled.

How much does Claude Sonnet 5.5 cost per coding session?

Identical token counts cost exactly what Sonnet 5 charged — the per-session win is whatever the token reduction delivers, which Anthropic caps its own claim at 30%. All prices verified against Anthropic’s official pricing docs on September 29, 2026:

Per million tokens	Sonnet 5.5	Sonnet 5	Opus 5.5	Fable 5.1
Input	$2.00	$2.00	$4.00	$10.00
Output	$10.00	$10.00	$20.00	$50.00
Cache read	$0.20	$0.20	$0.20	$0.25
5m cache write	$2.50	$2.50	$5.00	$12.50
1h cache write	$4.00	$4.00	$8.00	$20.00
Batch API	$1 / $5	$1 / $5	$2 / $10	$5 / $25

For the uncached 200K-input / 25K-output session shape we’ve used since the Opus 5 default cost analysis: Sonnet 5.5 runs $0.40 input + $0.25 output = $0.65, versus $1.30 on Opus 5.5 and $3.25 on Fable 5.1. If the token-efficiency claim holds at even half strength, real sessions land around $0.50-0.55.

The cache-heavy marathon shape from the Fable 5.1 analysis — 6M cache reads, 200K of 5-minute cache writes, 100K fresh input, 60K output:

Cost component	Sonnet 5.5	Opus 5.5	Fable 5.1
6M cache reads	$1.20	$1.20	$1.50
200K cache writes	$0.50	$1.00	$2.50
100K fresh input	$0.20	$0.40	$1.00
60K output	$0.60	$1.20	$3.00
Session total	$2.50	$3.80	$8.00

Note the cache-read row: since Opus 5.5 cut its cache reads to $0.20/M last week, the two models tie there — every dollar of Opus premium now comes from writes, fresh input, and output. At two marathon sessions per weekday, Sonnet 5.5 runs about $110/month against Opus 5.5’s $167 on identical token counts, before the fewer-tokens effect widens the gap.

Is Sonnet 5.5 actually as good as Opus 5.5 for coding?

On agentic terminal work, Anthropic’s own table says yes — it says Sonnet 5.5 is better. On the hardest problem sets, no. All scores below are vendor-run, published September 28, 2026, no independent verification yet:

Benchmark	Sonnet 5.5	Opus 5.5	Sonnet 5	Fable 5.1
Terminal-Bench 4.0	70.6%	66.4%	10.3%	55.8%
FrontierCode 1.1	46.2% (at max)	54.4%	42.4%	50.3%
CursorBench 4.0	55.5%	57.8%	34.1%	51.8%
GDPval-AA v2.1	1844	1846	1449	1735
OSWorld 2.1	80.1%	81.8%	57.0%	80.7%

Three caveats before you repeat these numbers:

• The Sonnet 5 column is not a typo, but it needs context. Terminal-Bench 4.0 and CursorBench 4.0 are newer, harder harness versions than the ones Sonnet 5’s launch numbers ran on — Sonnet 5 scored 63.2% on SWE-bench Pro in its day. A 10.3% here measures the old model against a benchmark generation it never targeted, which flatters the upgrade. The Sonnet 5.5 vs Opus 5.5 rows, run in the same table, are the reliable comparison.
• The FrontierCode score ran at max effort. That is the most expensive setting the model has, and it still trails Opus 5.5 by 8.2 points. Anthropic publishing the loss is to its credit — and it maps to a real recommendation: the hardest multi-file architecture work is still Opus territory.
• GDPval-AA is a 2-point tie. For document- and knowledge-heavy agent work, the launch table gives you no reason to pay double.

The strategic read: three weeks ago Anthropic’s Opus 5.5 table made Fable 5.1 look overpriced for coding. This table does the same thing to Opus 5.5 for everyday agentic sessions — the $2 model wins the benchmark most correlated with “run an agent in a terminal until the task is done.” What Opus keeps is the frontier-problem margin, which is exactly the work where a wrong answer costs more than the token bill.

How do you use Sonnet 5.5 in Claude Code?

On Claude Code v2.1.284 (released September 28, 2026) with Anthropic API billing, claude-sonnet-5-5 is already the default Sonnet — the changelog entry reads: “Added Claude Sonnet 5.5 (claude-sonnet-5-5), now the default Sonnet model on the Anthropic API.” Verified against the official Claude Code changelog on September 29, 2026.

Flat-rate subscribers should know a second change from the same week: v2.1.280 flipped the default model on Pro and Team Standard plans from Sonnet to Opus, matching Max and Enterprise. So out of the box, plan users get Opus 5.5 and API-billed users riding the sonnet alias get Sonnet 5.5. Pin what you actually want:

/model              # opens the picker — Enter saves as default
/model sonnet       # latest Sonnet (5.5 on v2.1.284+, API billing)
/model opusplan     # Opus for plan mode, Sonnet for execution — the budget hybrid

For API-billed Claude Code users the opusplan split just got more attractive: planning happens on the model that wins FrontierCode, execution happens on the model that wins Terminal-Bench, and the execution phase — where most tokens burn — bills at half price.

How do you use Sonnet 5.5 in Cursor and GitHub Copilot?

Cursor listed Claude Sonnet 5.5 on launch day — enable it under Settings → Models, no BYOK required. Per Cursor’s model docs (verified September 29, 2026 via search cross-reference; cursor.com blocks our direct fetch): it scores 55.5% on CursorBench at max effort, second only to Opus 5.5 on Cursor’s own leaderboard, keeps Sonnet 5’s per-token price in the usage pool, supports thinking, and carries the full 1M context. For Cursor Pro ($20/month, verified September 21) users, a Sonnet that benchmarks within 2.3 points of Opus 5.5 stretches included usage roughly twice as far.

GitHub Copilot shipped Sonnet 5.5 the same day to Pro, Pro+, Max, Business, and Enterprise plans via the model picker, with a gradual rollout and admin enablement through the model policy on Business/Enterprise. GitHub’s changelog adds a detail worth quoting: in its early testing, Sonnet 5.5 “matched Claude Sonnet 5 on coding tasks while using significantly fewer steps, tokens, and tool calls.” Under Copilot’s usage-based billing at provider list pricing, fewer tokens is the whole ballgame — same subscription, more work per premium request.

Can Cline select Sonnet 5.5 yet?

Probably not from the built-in catalog — and we can’t fully re-verify today, so here is exactly what we know. When we checked Cline’s source for the Opus 5.5 article on September 23, the then-current v4.1.20 catalog predated the entire 5.5 generation, and Cline’s Anthropic provider only accepts catalog models — free-text IDs are reserved for the openai-compatible, ollama, lmstudio, litellm, and vertex providers. Cline’s changelog was unreachable from our verification environment today, so treat the catalog gap as “last confirmed September 23” and check your extension’s model list before assuming either way.

Two routes that work regardless:

1. OpenRouter provider. Cline fetches OpenRouter’s model list live, and OpenRouter lists anthropic/claude-sonnet-5.5 served by five providers (verified September 29, 2026). Add an OpenRouter key, refresh, select it — you pay a routing margin but run the new model today.
2. Vertex provider with a typed ID. Google Cloud users can enter claude-sonnet-5-5 directly; Vertex accepts custom model IDs.

And the standing warning from the Opus 5.5 article still applies: Cline’s unpinned Anthropic-provider default resolves to Fable 5.1 at $10/$50 — a model that now loses to a $2/$10 Sonnet on Anthropic’s own Terminal-Bench row. If you have never pinned a model in Cline’s provider settings, you are paying 5x per output token for lower vendor-run agentic scores. Pin something today.

Five breaking changes if you call the API directly

BYOK harnesses moving from claude-sonnet-5 to claude-sonnet-5-5 hit five breaking changes, verified against Anthropic’s model docs on September 29, 2026:

1. thinking: {"type": "disabled"} returns a 400. The new off-switch is {"type": "between_tools"}, which disables up-front thinking, works only at high effort or below, and accepts no other field alongside it.
2. Forced tool use is gone. tool_choice: {"type": "any"} and {"type": "tool", ...} return a 400. Use auto plus a prompt instruction, strict: true for schema-valid arguments, or structured outputs.
3. Sampling parameters are rejected. Setting temperature, top_p, or top_k to any non-default value returns a 400. This is the sleeper: plenty of Cline and custom-harness configs pin temperature: 0 for determinism. Delete the parameter.
4. Thinking blocks are tied to the model and the conversation (“preserved thinking”). Editing earlier turns invalidates them; accounts created on or after August 31, 2026 get a hard 400 on edited history. Make your harness append-only.
5. The older computer_20251124 computer-use tool is rejected on the Claude API and Google Cloud — only computer_toolset_20260801 works. The advisor tool also rejects Opus 4.8, Opus 4.7, and Sonnet 5 as advisors.

One non-breaking change that confuses UIs: text between tool calls now arrives as thinking blocks that are empty at the default display setting. A harness that streams that text as progress updates goes silent mid-loop until it sets a display value that returns it — or switches to between_tools.

When should you NOT switch to Sonnet 5.5?

Four honest cases:

• Your hardest work is the work. An 8.2-point FrontierCode gap at max effort is not noise. Multi-file architectural refactors, gnarly concurrency bugs, work where a subtle miss costs hours — run those sessions on Opus 5.5 and let Sonnet 5.5 handle the rest.
• Your harness sets temperature or uses forced tool_choice. Those are hard 400s now, not warnings. Sonnet 5 remains available at the same price while you migrate.
• You edit conversation history programmatically on a post-August-31 API account. Preserved thinking will reject your requests until the harness goes append-only.
• You need independent benchmarks before a team standardizes. Every score above is vendor-run and one day old. History says a few points of drift from launch numbers is normal once public leaderboards post.

If your real constraint is that any metered API bill is a problem: a used RTX 3090 running Qwen3-Coder locally charges $0/month in tokens (hardware guide on our sister site), and you can trial the workload on a rented 3090 from about $0.07/hr on Vast.ai before buying hardware. Open-weight quality sits well below Sonnet 5.5, but no meter runs — see the FOSS harness comparison on aifoss.dev for which local-first tool handles it best.

Which model should you actually run now?

Prices as of September 2026, all verified in the tables above:

Your situation	Run this	Cost	Where
API-billed agentic coding, most sessions	Sonnet 5.5	$2 / $10 per MTok	Claude API
The hardest 20% — architecture, deep debugging	Opus 5.5	$4 / $20 per MTok	Claude API
Claude Code, API billing	/model sonnet or opusplan (Sonnet executes, Opus plans)	Session math above	Claude Code changelog
Cursor, no BYOK	Sonnet 5.5 in the picker	Plan usage at $2/$10 rates	cursor.com
Copilot Pro/Business	Sonnet 5.5 via model picker (admin-enabled on Business)	Included; usage-based at list price	GitHub Copilot
Cline today	OpenRouter anthropic/claude-sonnet-5.5, or pin Sonnet 5 until the catalog updates	$2/$10 + routing margin	cline.bot
Still running Fable 5.1 or Sonnet 5 for coding	Switch — 5.5 wins on price-per-score either way	—	Claude API
Zero-meter / privacy-bound	Local Qwen3-Coder on a 24GB GPU	$0/mo tokens; rent from ~$0.07/hr first	Vast.ai

FAQ

Is Claude Sonnet 5.5 cheaper than Sonnet 5? Per token, no — identical on every line: $2/$10 base, $2.50 5-minute cache writes, $0.20 cache reads, $1/$5 on the Batch API (verified September 29, 2026). Anthropic’s “up to 30% cheaper per task” claim rests entirely on the model using fewer tokens for the same work, which its launch customers (Balyasny: 121K vs 497K tokens per answer; Box: 12% fewer tokens) corroborate directionally.

Does Sonnet 5.5 really beat Opus 5.5 at coding? On Terminal-Bench 4.0, yes — 70.6% vs 66.4%, vendor-run. Opus 5.5 wins FrontierCode (54.4% vs 46.2%), CursorBench (57.8% vs 55.5%), and OSWorld (81.8% vs 80.1%). Read it as: Sonnet 5.5 for terminal-agent grunt work, Opus 5.5 for the hardest problems, and a coin flip on knowledge work (GDPval: 1846 vs 1844).

Why does Sonnet 5 score only 10.3% on Terminal-Bench in the launch table? Terminal-Bench 4.0 is a newer, harder harness generation than the benchmarks Sonnet 5 launched against in June 2026 (it scored 63.2% on SWE-bench Pro then). Old model, new benchmark — the gap is real but the framing flatters the upgrade. Compare Sonnet 5.5 against Opus 5.5 instead; both were tuned for this generation.

Is Sonnet 5.5 in Claude Code by default? On v2.1.284+ with Anthropic API billing, the sonnet alias resolves to Sonnet 5.5. Pro and Team Standard plan users default to Opus since v2.1.280 — run /model sonnet to switch, or opusplan to split planning and execution.

Did Sonnet pricing go up to $3/$15? No. That increase was scheduled for September 1, 2026 when Sonnet 5’s $2/$10 was announced as introductory, and Anthropic’s pricing page now confirms it was cancelled — $2/$10 is the standard price for Sonnet 5 and 5.5 alike. Ignore any source still quoting $3/$15.

Sources

• Introducing Claude Sonnet 5.5 — Anthropic announcement (benchmarks, speed and cost claims)
• Claude Sonnet 5.5 overview — Claude Platform docs (specs, pricing, breaking changes)
• Pricing — Claude Platform docs (full price list, Sonnet 5 introductory-pricing note)
• Claude Code changelog — v2.1.284 default Sonnet change, v2.1.280 plan default change
• Claude Sonnet 5.5 — Cursor docs (model picker, CursorBench score)
• Claude Sonnet 5.5 in GitHub Copilot — GitHub Changelog (plans, rollout, token-efficiency note)
• Claude Sonnet 5.5 — OpenRouter model listing (Cline workaround route)
• Anthropic releases Sonnet 5.5 — TechCrunch coverage

Last verified September 29, 2026. Model pricing, benchmark leaderboards, and tool integrations change frequently; check the official pages above before committing a team or a budget.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe