2026-09-23 🧭 Daily News

Claude Opus 5.5 Launches: Fable-Level Performance at 40% Lower Cost, Plus Code v2.1.280

🧭 Claude Opus 5.5: Fable-Level Performance at 40% Lower Cost, With Always-On Adaptive Thinking

Anthropic has launched Claude Opus 5.5, the first model in the new 5.5 efficiency tier. It delivers performance on par with Claude Fable 5.1 on most workloads — including coding, multi-step reasoning, and document analysis — while costing 40% less per token than Opus 5 and generating output more than 30% faster. Sonnet 5.5 and Haiku 5.5 will follow within weeks, completing the efficiency-tier trio.

Key specs

Context window: 1,000,000 tokens (1M) — same as Fable 5.1
Max output: 128,000 tokens per response
Adaptive thinking: Always-on; no extra flag or budget parameter required — the model allocates reasoning steps automatically based on task difficulty
Inline tools in mid-conversation system messages: Beta — system messages can now inject or remove tool definitions between turns without rebuilding the context
Pricing: $4 / $20 per million input/output tokens; cache reads $0.20/MTok (60% cheaper than Opus 5 cache reads)
Availability: Anthropic API, AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure AI

No more 5-hour usage caps

Alongside the model launch, Anthropic is removing the 5-hour hard usage cap that applied to Pro, Max, Team, and seat-based Enterprise subscribers when running Opus-tier models. A new rate-limit reset feature lets users manually reset their per-window counter — useful for burst workflows that genuinely need more in a single session. Heavy use will still be subject to soft rate limits at sustained load, but the 5-hour wall is gone.

Developer decision guide: Opus 5 vs Opus 5.5

If your pipeline is latency-sensitive or runs at scale, Opus 5.5 is an immediate drop-in replacement — same capability tier, better throughput, lower bill. The key caveat is the always-on adaptive thinking: unlike Opus 5 where you controlled extended-thinking budgets explicitly, Opus 5.5 allocates reasoning time internally. In practice this is transparent for most tasks, but if you were deliberately suppressing thinking tokens to cut cost, you lose that knob. Teams doing tightly metered token accounting should benchmark a representative sample before migrating production workloads. The inline-tools-in-system-messages beta is worth enabling immediately if your agent architecture needs to compose tool sets dynamically — it eliminates the need to restart context when the active tool set changes mid-session.

# Switch an existing API call to Opus 5.5
import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5-5",          # new model ID
    max_tokens=8192,
    messages=[{"role": "user", "content": "Analyse this codebase and suggest refactors."}]
)
# Adaptive thinking is automatic — no `thinking` param needed
print(response.content[0].text)

# Beta: inject tools mid-conversation via system message
# Pass betas=["inline-tools-2026-09-22"] in the request headers
# to enable dynamic tool injection between turns.

⭐⭐⭐ anthropic.com

🧭 Claude Code v2.1.280: Opus 5.5 as Default, MCP Description Caps, and 113 Fixes

Claude Code v2.1.280 (released September 22) picks up Opus 5.5 as the new default Opus-tier model and ships a dense round of 113 changes across model defaults, MCP controls, UI/UX, session reliability, and integrations. The headline developer-facing additions are a new environment variable for MCP description length, a fix for prompt-cache misses on live model switches, and a correction to how resumed fork subagents rebuild their tool lists.

Changes worth acting on

Opus 5.5 is now the default Opus model — if you have "model": "opus" in your .claude/settings.json, you're already on 5.5 after upgrading. Pin to claude-opus-5 explicitly if you need the prior model.
CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH — new env var overrides the 2,048-character cap on MCP tool descriptions and server instructions. Set a larger value if your MCP tools have verbose schemas that were being silently truncated.
PermissionRequest hook change — agent-type hooks no longer fire in the PermissionRequest hook; an error message now points you to command or http hooks instead. If you have custom permission scripts, check whether they use the agent hook type.
Model-switch prompt-cache fix — switching the active model from a host app while Claude is processing a message no longer causes a cache miss on the next prompt. Long agentic sessions with dynamic model switching will see meaningfully lower token costs.
Fork subagent tool list fix — resumed fork subagents now re-send the original tool list rather than rebuilding it, preventing tool-schema divergence in long-running parallel workflows.

Check your MCP tool description lengths

The 2,048-character default cap has been silently truncating MCP tool descriptions longer than that limit since MCP support launched. If your tools have rich JSON schemas or long description fields, they may have been sending incomplete metadata to Claude, which can cause subtle misuse of tool parameters. After upgrading to v2.1.280, set CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH=8192 (or your actual max), run claude /mcp list, and compare tool descriptions against your MCP server definitions. Any mismatch before this version warrants a re-test of tool-calling accuracy on those tools.

# Upgrade to v2.1.280
npm install -g @anthropic-ai/claude-code@latest

# Confirm the default Opus model
claude /config
# → Default model (opus tier): claude-opus-5-5

# Override MCP description cap (e.g. allow up to 8 KB)
export CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH=8192
claude /mcp list   # descriptions now shown in full

# Pin to previous Opus if needed
# In .claude/settings.json:
# { "model": "claude-opus-5" }

⭐⭐⭐ github.com/anthropics/claude-code

Source trust ratings ⭐⭐⭐ Official Anthropic · ⭐⭐ Established press · ⭐ Community / research