Claude Sonnet 5.5: Everyday Coding Ladder Beside Opus 5.5 — Cost Per Task, Not Another Flagship Rematch

Last night’s Naive-N0.5-Flash note was a decode-cost ticket — zero full-attention layers at native 1M. Yesterday morning’s OpenShell was runtime policy. The live Claude Opus 5.5 post closed with an asterisk: Sonnet and Haiku 5.5 were still “coming weeks,” so do not assume the whole Claude ladder prices like Opus. On September 28, 2026, Anthropic shipped Claude Sonnet 5.5. This morning is that mid-tier board — when to route Sonnet vs Opus after both 5.5 ships exist, not another Opus rematch.

Primary sources: Anthropic launch, Sonnet 5.5 overview docs. Optional secondary: The Decoder. Specs and benches below are Anthropic’s. Treat cost-per-task charts, generation-speed claims, and eval rows as vendor claims.

Spec sheet (named numbers)

Opus 5.5 stays $4/$20. Sonnet 5.5 stays $2/$10. The claim is not a cheaper sticker than Sonnet 5 — it is fewer tokens per finished task and faster generation at the same list, with vendor benches near Opus 5.5 on several agentic evals at Max.

What actually changed for people who route agents

The durable change is the Sonnet ladder beside a live Opus 5.5, not a new FLOPS sticker.

Anthropic’s case: same $2/$10 as Sonnet 5, but up to 30% lower cost per task (fewer tokens) plus 30%+ faster generation. At Low/Medium effort, Anthropic’s charts say Sonnet 5.5 often beats Sonnet 5’s best for roughly 1/10 the cost per task — vendor $/task until you reproduce it.

Effort defaults: Claude API high, apps Medium — different from Opus 5.5’s medium-default in the Opus post. Log the effort you send.

Safeguards: first Sonnet with Opus-class cyber safeguards and fallbacks, plus distillation classifiers. Biology stays at Sonnet 5 level; preserved thinking stays on. When cyber fires, the request can fall back to another Claude model — log model id per step.

Migration: thinking off on Sonnet 5 → set between_tools before cutover. Non-default temperature / top_p / top_k → 400. Fix sampling in the client first.

Vendor benches (Anthropic’s table)

Footnotes (Anthropic): (1) Opus Terminal-Bench at Xhigh. (2) Sonnet Max can score lower than Xhigh on FrontierCode (code-review skill / timeout / out-of-scope edits). (3) AA ran GDPval/Briefcase on pre-release with a structured-outputs bug (fixed; may understate). (4) GPT-6 Sol image-understanding bug recently fixed; AA scores may lag.

Anthropic’s read: Sonnet 5.5 at Max is near Opus 5.5 on several evals, but Opus remains clearly stronger at complex open-ended judgment. SpaceXAI on CursorBench: “frontier-level… 55.5%, second only to Opus 5.5.”

Fine print: Terminal-Bench 70.6% vs Sonnet 5’s 10.3% needs a your-harness re-run; FrontierCode Max vs Xhigh can invert; GDPval/Briefcase sit within a few Elo of Opus (AA understatement risk); small Opus gaps elsewhere are unresolved until you bake off.

Where it sits next to Naive, OpenShell, JEV, Holo4, Opus, and Sol

Naive = decode architecture. OpenShell = process policy. Opus = hard unattended / open-ended judgment. Sonnet 5.5 = where most coding agent loops should land once the 5.5 mid-tier exists. Keep Sol as the OpenAI $2/$10 peer; do not collapse Sonnet into Opus or Naive.

Caveats I would write into any Sonnet 5.5 ticket

1. Vendor harness ≠ your fleet. Re-run Terminal-Bench / FrontierCode / CursorBench / your repo before moving default traffic off Sonnet 5 or Sol.
2. Same list as Sonnet 5. The win is tokens-per-task and speed — measure dollars per successful merge.
3. Near-Opus ≠ Opus. Route complex open-ended judgment to Opus 5.5.
4. Effort defaults differ (API high vs apps Medium). Pin effort in the client.
5. Cyber fallbacks. Log which model finished the turn.
6. Migration 400s. Fix between_tools and non-default sampling before cutover.
7. AA / Sol footnotes. Pre-release structured-outputs bug and Sol image bug can move Elo rows.
8. Not Naive / OpenShell / Opus rematch. This is the Sonnet cost-per-task ladder.

How I would evaluate if I routed everyday coding traffic

1. Twenty-task bake-off — migrations, overnight unattended jobs, one citation-checked pack. Run Sonnet 5.5, Sonnet 5, Opus 5.5, and GPT-6 Sol at Medium, high, and Max/Xhigh where the API allows.
2. Cost per success — wall time, tokens, dollars per successful merge, rework minutes. Prefer Sonnet 5.5 over Sonnet 5 only when both move.
3. Route split — Opus 5.5 on open-ended judgment and hardest unattended loops; everyday coding on Sonnet 5.5 if the bake-off holds.
4. Safeguard logging — record model id per step when cyber fallbacks fire.
5. Keep boards separate — Naive for decode-cost, OpenShell for policy, JEV for System 1, Holo4 for computer-use, Opus for flagship judgment, Sol for the OpenAI $2/$10 peer. Sonnet 5.5 is the Claude everyday coding board.

Bottom line

On September 28, 2026, Anthropic shipped Claude Sonnet 5.5 (claude-sonnet-5-5) at $2 / $10 with $0.20 cache reads — same list as Sonnet 5 — claiming up to 30% lower cost per task and 30%+ faster generation, with vendor benches near Opus 5.5 on several Max evals while Opus remains the open-ended judgment SKU. Durable facts: rate card across Claude API / Bedrock / Google Cloud / Foundry, adaptive thinking (high on API), first Sonnet with Opus-class cyber safeguards, and migration breakers (between_tools; non-default sampling → 400).

That closes the mid-tier asterisk the Opus post left open. Not a Naive night, not an OpenShell night, and not permission to retire Opus without a judgment-lane bake-off. Measure cost per successful merge, then move everyday coding traffic.

Create-only. Specs and benches from Anthropic’s Sep 28 launch page and Sonnet 5.5 docs. Update when Haiku 5.5 ships, when a named third party reproduces Terminal-Bench / FrontierCode / CursorBench under a common harness, or when live Batch 300K availability is confirmed.

Sources: Anthropic · Sonnet 5.5 docs · The Decoder.

Anpoo Sivanadi, Staff Software Engineer