← Back to Payloads
AI News2026-09-28

Claude Sonnet 5.5 Just Became Claude Code's Default Sonnet. 1M Context, $2/$10, $0.20/MTok Cache Reads, and Five Breaking API Changes.

Anthropic pushed Claude Sonnet 5.5 to GA on September 28 and made it the default Sonnet model in Claude Code v2.1.284 the same day. List price stays at $2/$10 per MTok with $0.20/MTok cache reads (unchanged from Sonnet 5), but the 1M context window, the always-on adaptive thinking, and five breaking API changes make the model upgrade consequential for agent stacks that wrote prompts against Sonnet 5. Documentation indicates code that turns thinking off explicitly, or relies on forced tool use, will return 400s on Sonnet 5.5.

What happened

On 2026-09-28 at 18:02 UTC, Anthropic released Claude Code v2.1.284, and the headline change is the model default flip: Claude Sonnet 5.5 (claude-sonnet-5-5) is now the default Sonnet model on the Anthropic API. The release notes state verbatim: "Added Claude Sonnet 5.5 (claude-sonnet-5-5), now the default Sonnet model on the Anthropic API — 1M context, $2/$10 per Mtok with $0.20/Mtok cache reads."

Anthropic's platform release notes confirm a coordinated model launch: "We've launched Claude Sonnet 5.5 (claude-sonnet-5-5). It's available on the Claude API, Claude in Amazon Bedrock…" alongside a Sonnet 5.5 prompt-engineering reference page and a Sonnet 5.5 system-prompts page. The same announcement cycle that surfaced Sonnet 5.5 includes the breaking-change documentation Anthropic wrote specifically to flag the upgrade path.

Claude Code v2.1.284 is a substantial release beyond the model flip: more than 60 distinct changes, including dollar amounts in /usage and the status line, a Yes, but ask again next time answer in auto mode, /mcp reconnect all, auth: { google: {} } for OTLP telemetry on Google Cloud, and private_key_jwt client authentication between the Claude apps gateway and identity providers. Three release categories matter for agent-stack builders: model+pricing, breaking API changes, and gateway/auth expansion.

What actually changed

Model default. Sonnet 5.5 replaces Sonnet 5 as the default Sonnet model on the Claude API. The Sonnet 5 model itself is not removed, but anyone who has not pinned model in their Claude Code settings or in API request bodies is now talking to Sonnet 5.5 by default.

Pricing. Documentation indicates the list price is unchanged from Sonnet 5 at $2 input / $10 output per MTok. Cache reads remain at $0.20 per MTok (the standard 0.1x multiplier on the $2 input price). The release notes and pricing-page position are consistent on this point.

Context and thinking. Sonnet 5.5 ships with a 1M-token context window and always-on adaptive thinking. The model is documented on Anthropic's platform release notes as available on Claude API, Amazon Bedrock, and other first-party surfaces; cross-cloud parity is not separately verified for Vertex AI or Microsoft Foundry in this sweep.

Breaking changes. The Anthropic release notes for Sonnet 5.5 state: "Code written for Claude Sonnet 5 can break on Claude Sonnet 5.5 in five ways." The most consequential one called out in the platform documentation is the thinking-disable path: "To turn off up-front thinking, send thinking: {"type": "between_tools"} instead of "disabled", at high effort or below." Forced tool use and thinking-block model-binding changes also recur from the Opus 5.5 breaking-change pattern published with that launch — documentation indicates the same shape applies.

Claude Code operational changes. Three classes of changes matter for production deployments. First, the gateway spend-limit display in /usage and the rate_limits.spend_limit status-line field now expose used_usd, limit_usd, and period. Second, claude mcp add no longer reports success when managed settings restrict MCP servers to plugins — it now refuses with an actionable message. Third, ANTHROPIC_FOUNDRY_RESOURCE is now validated; a value that is not a plain resource name is refused before the request is sent, closing a host-injection vector on Microsoft Foundry.

VS Code and Cloud sessions. The release ships parallel improvements to the VS Code extension (Ultracode toggle under Effort slider, Message Timestamps setting, Reload Claude saving before restart) and Cloud sessions (offline detection on routine Edit and Duplicate controls).

Why developers and founders should care

The model flip is silent. Existing Claude Code users who never set model will see different behaviour tomorrow morning: better reasoning at the same list price, but a different response shape when prompts request disabled thinking, a different error when prompts force a specific tool, and different cache-key semantics that documentation indicates may invalidate warm caches during the cutover. For API users, the same silent default applies — request bodies without an explicit model field are now routed to Sonnet 5.5.

The cache-read economics stay flat against Sonnet 5, so there is no immediate budget win or loss for caches that already ran on Sonnet 5. The headline story is reliability: a 1M context window with adaptive thinking changes long-context agent feasibility in the same price band, and the breaking-change flag means prompt migration matters before the next agent rollout.

For gateway operators and MCP-heavy teams, the operational changes are more material than the model itself. The dollar display in /usage and the status line make per-team and per-session spend visible without a separate dashboard. The MCP rejection path closes a "configuration that silently never loaded" failure mode. The Foundry resource validation closes an injection surface on Azure deployments.

Evidence and verification

  • Claude Code v2.1.284 GitHub release tag, atom feed updated 2026-09-28T18:01:46Z; release notes inspected verbatim (model default, pricing, security, gateway, VS Code, Cloud sessions, Claude Tag, Code Review).
  • Anthropic platform release notes (overview page, fetched 2026-09-28 20:08 UTC): "We've launched Claude Sonnet 5.5 (claude-sonnet-5-5)." model-launch language and Sonnet 5.5 prompt-engineering reference page path both present.
  • Anthropic pricing page: Sonnet 5.5 row visible above Sonnet 5 row, input $2 / output $10, cache-read footnote unchanged from Sonnet 5 at 0.1x input = $0.20/MTok.
  • Breaking-change summary: Anthropic platform release notes for Sonnet 5.5 state "Code written for Claude Sonnet 5 can break on Claude Sonnet 5.5 in five ways." Specific case (thinking disable) quoted verbatim from the platform documentation.

Cost, risk, and limitations

Cost. No list-price change from Sonnet 5. Cache-read economics unchanged. Spend display added in /usage makes the per-session cost visible but does not change it.

Risk. Three classes of breakage to watch. First, prompts that send thinking: {"type": "disabled"} at high effort or below will fail or behave differently on Sonnet 5.5; switch to {"type": "between_tools"}. Second, forced tool use on a specific tool returns a different error class — documentation indicates requests that rely on the older forced-tool-use contract will see 400s. Third, thinking-block model binding means cached sessions created against Sonnet 5 may not transfer cleanly to Sonnet 5.5; treat warm caches as cold for the first few hours of cutover.

Limitations. This is a documentation-surfacing report. No firsthand API request was sent. Cross-cloud parity for Vertex AI and Microsoft Foundry is not separately verified in this sweep. Pricing-page text was retrieved from the live page; a future pricing update could shift the cache-read multiplier.

Mr. Technology verdict

This is a quiet model upgrade with quiet operational upgrades underneath. The model flip is consequential because it is silent — anyone not pinning a model is now on Sonnet 5.5 — but the headline cost story is "unchanged list price, $0.20 cache reads, 1M context, always-on adaptive thinking." The breaking-change flag is the part to take seriously if your prompt library turns thinking off explicitly, relies on forced tool use, or assumes specific cache-key behaviour across sessions. For agent stacks that already run on Sonnet 5 with the default behaviour, the cutover is free. For prompt libraries that override defaults, schedule a migration pass before the next agent rollout.

Recommended action

  • Pin model: claude-sonnet-5 explicitly in your Claude Code settings and API request bodies if you need to stay on Sonnet 5 during migration; otherwise, accept the silent default flip to Sonnet 5.5.
  • Audit prompt libraries for thinking: {"type": "disabled"} and forced-tool-use patterns. Replace disabled with between_tools at high effort or below where the existing prompt relied on disabling thinking. Treat forced tool use as a breaking-change surface; plan to migrate to the new contract before rolling out new agent prompts.
  • For Gateway operators: verify rate_limits.spend_limit now exposes used_usd, limit_usd, and period; wire those into your spend dashboards. Verify claude mcp add is now refusing plugins-restricted configurations with a usable message rather than silently accepting them.
  • For Microsoft Foundry deployments: confirm ANTHROPIC_FOUNDRY_RESOURCE is set to a plain resource name before deploying v2.1.284; the unvalidated interpolation path is closed in this release.

Sources

  • Claude Code v2.1.284 release notes — https://github.com/anthropics/claude-code/releases/tag/v2.1.284 (verified 2026-09-28 20:08 UTC)
  • Anthropic platform release notes (Sonnet 5.5 launch) — https://platform.claude.com/docs/en/release-notes/overview (verified 2026-09-28 20:08 UTC)
  • Anthropic Sonnet 5.5 prompt engineering — https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5-5 (referenced 2026-09-28 20:08 UTC)
  • Anthropic pricing — https://platform.claude.com/docs/en/about-claude/pricing (verified 2026-09-28 20:08 UTC)

Documentation comparison; no firsthand API call. Pricing, model ID, and breaking-change summary sourced verbatim from Anthropic platform documentation and Claude Code v2.1.284 release notes.

Related Dispatches