claude/for
Search
Subscribe
Plans & ModelsUpdated

Claude Sonnet 5.5 (Sep 28, 2026): $2/$10, Platform breaks, Claude Code 2.1.284 default Sonnet

Arthur Teboul
Arthur TeboulEditor, claude/for
UPDATED OCT 08 · 5 MIN

Sep 28 launch: Sonnet 5.5 at $2/$10, cache reads $0.10 since Oct 7; five Platform migration breaks; Claude Code 2.1.284 default Sonnet + auto-mode start.

On Sep 28, 2026, Anthropic launched Claude Sonnet 5.5 (claude-sonnet-5-5): same list price $2 / $10 per MTok. Cache reads launched at $0.20 and were cut to $0.10 on Oct 7, 2026 (Platform RN). Five Platform breaks vs Sonnet 5 (thinking between_tools not disabled; tool_choice any/tool → 400; account-/conversation-bound thinking blocks; computer_20251124 rejected on API/GCP; advisor rejects Opus 4.8 / 4.7 / Sonnet 5). Claude Code 2.1.284 makes it the default Sonnet and starts interactive/VS Code sessions in auto mode when no permission mode is set. See also Claude Opus 5.5, Claude Haiku 5.5 (launched Oct 7), and auto mode.

Pricing and role

Per Anthropic’s pricing docs, checked Oct 8, 2026 (vs Opus 5.5):

Per 1M tokensSonnet 5.5Opus 5.5
Input$2$4
Output$10$20
Cache reads$0.10$0.20
Cache writes$2.50$5

Sonnet 5.5 cache reads cost 0.05x base input since Oct 7, 2026 (they launched at $0.20, the 0.1x rate); cache writes and all other prices are unchanged, per the Platform RN.

Same list price as Sonnet 5. Anthropic claims outputs 30%+ faster and up to 30% less cost per task vs Sonnet 5 because it typically needs fewer tokens — attributed product-card language, not a CF measurement. Positioning: faster, lower-cost complement to Opus 5.5 for well-scoped everyday work, bugs, and polished docs/slides/spreadsheets; design eye. Haiku 5.5 launched on Oct 7, 2026: see Claude Haiku 5.5. Anthropic’s product card reports Terminal-Bench 4.0 70.6% vs Sonnet 5 10.3%. Model id on Platform: claude-sonnet-5-5 (API, Bedrock, Platform-on-AWS, GCP, Foundry).

The Claude moves worth stealing, by email

Workflows tested on real jobs, sent only when they are worth it. Free.

Platform migration breaks

Platform RN September 28, 2026 + What’s new in Sonnet 5.5 — five breaks vs Sonnet 5 code:

  1. Thinking off. Send thinking: {"type": "between_tools"} instead of "disabled" (at high effort or below). "disabled" → 400.

  2. Forced tools. tool_choice types any / tool → 400. Use auto (with strict tool use) or none.

  3. Thinking blocks. Tied to the model and conversation; Sonnet 5.5 blocks are also account-/conversation-bound (Preserved thinking). Another account → block dropped, request still succeeds.

  4. Computer use. On Claude API and Google Cloud, earlier computer_20251124 is not accepted — migrate to the current computer toolset (computer_toolset_20260801 per what’s-new). On Amazon Bedrock, what’s-new states Sonnet 5.5 still accepts computer_20251124.

  5. Advisor. Rejects Opus 4.8 / Opus 4.7 / Sonnet 5 as advisors for a Sonnet 5.5 executor.

Keep migration thin — follow Anthropic’s what’s-new + migration guide; no full agent-loop rewrite here.

Claude Code 2.1.284

Raw CHANGELOG head was ## 2.1.284 at launch. Key lines:

The rest of 2.1.284 (UI, hooks, and VS Code changes) stays in the CHANGELOG.

Upgrade readout

OperatorAction
API / GCP integratorsMigrate thinking-off → between_tools; drop forced tool_choice; move computer use off computer_20251124; update advisor pairings
Bedrock computer-usecomputer_20251124 still OK on Sonnet 5.5 (what’s-new)
Claude CodeUpgrade to ≥ 2.1.284; check /model and permission mode after Sonnet default + auto-mode default
Opus usersSee Claude Opus 5.5

Related operator guides: max effort · rate limits.

Sources (checked 2026-09-29; pricing rechecked 2026-10-08)

The Claude moves worth stealing, by email

Workflows tested on real jobs, sent only when they are worth it. Free.