speed parameter landing natively in the SDK signals that fast mode has graduated from something you jury-rig with beta headers to a first-class, documented, multi-language feature with rate-limit headers, fallback patterns, and a usage.speed field you can observe in production. Meanwhile, pydantic-ai shipped MetaProvider the day after Meta's Muse Spark Open landed — the ecosystem is catching up with model releases faster than ever. Two behavior-change warnings deserve attention before anyone upgrades blindly: the OTel SDK integration now auto-emits spans if you have a global tracer provider, and switching between speed: "fast" and standard invalidates your prompt cache.speed parameter [MEDIUM]: fast mode is now a first-class Messages API parameter with code examples in 8 languages; pricing is 2x standard; cache gotcha: fast/standard requests don't share cached prefixes.Anthropic(open_telemetry=False) before upgrading in production.npm install -g billion-context
ANTHROPIC_BASE_URL="http://localhost:8787/bili/https://api.anthropic.com" claude
bili plugin install claude
No breaking changes this period.
No model releases in the last 24 hours.
fast-mode-2026-02-01 beta header in HTTP requests — there was no SDK-level parameter. As of Python SDK v1.15.0 and TypeScript SDK v0.135.0 (both Oct 15), speed="fast" is a first-class top-level parameter on messages.create(), and the response's usage object now includes a speed field ("fast" or "standard") that lets you verify which tier actually ran.speed="fast" to Messages API across 8 SDK languages; Opus 5.5 needs no beta header, Opus 5 and 4.8 still do; pricing is 2x standard ($8/$40 per MTok for Opus 5.5); prompt cache is not shared between fast and standard speed requests.speed="fast" to your messages.create() call — no beta header needed:response = client.messages.create(
model="claude-opus-5-5",
max_tokens=4096,
speed="fast",
messages=[{"role": "user", "content": "..."}],
)
print(response.usage.speed) # "fast" or "standard"
For Claude Opus 5 and Opus 4.8, still send anthropic-beta: fast-mode-2026-02-01 alongside the speed parameter. Critical cache gotcha: fast and standard speed requests do NOT share prompt cache prefixes — if you have a large cached system prompt and switch between speed: "fast" and standard (e.g., in a fallback), you'll take a cache miss. Design your fallback pattern with this in mind. Rate limits for fast mode are separate from standard Opus rate limits and have their own set of response headers (anthropic-fast-input-tokens-limit, etc.). Fast mode is NOT available on the Batch API, Priority Tier, or Claude Platform on AWS.
speed="fast" to existing calls; header approach still works). Moderate if you also update fallback logic for cache-miss considerations.anthropic.messages.create spans to be emitted and forwarded to your observability backend — no configuration required. The opt-out is explicit.Anthropic(open_telemetry=False) or ANTHROPIC_OPEN_TELEMETRY=false; supported in Claude API, Bedrock, Vertex, Google Cloud, Foundry, and AWS SDKs.client = Anthropic(open_telemetry=False)
const client = new Anthropic({ openTelemetry: false });
If you DO want traces, the default content mode is metadata_only (no prompts or responses in spans) — to include content, set traces.content_mode = "content". The same SDK update also adds execution_identity to Managed Agents and telemetry destination management to the Admin API. This SDK change covers all cloud variants: Bedrock, Vertex, Google Cloud, Foundry, and AWS SDKs all received the same OTel update in the same batch.
open_telemetry=False if you don't want spans; review your OTel pipeline if you do).speed: "fast" in your Foundry SDK calls; not available for Anthropic-hosted Foundry or Claude Platform on AWS.speed="fast" parameter via the Foundry SDK variant (foundry-sdk-v0.6.0 also shipped with OTel support on the same date). Check platform.claude.com/docs/en/build-with-claude/claude-in-microsoft-foundry for the Foundry-specific SDK setup.speed="fast" to existing Foundry SDK calls, no additional config needed for Azure-hosted deployments).arXiv RSS feeds (cs.AI, cs.CL, cs.LG, cs.CV) all returned DNS failures this scan. HuggingFace Papers DNS failure. No qualifying research papers surfaced in the October 15–16 window. Section omitted.
MetaProvider that connects pydantic-ai agents to Meta Muse models via the Responses API. Separately, a new ApprovalReviewer harness lets a model approve low-risk tool calls before a human reviewer sees them — a step toward tiered human-in-the-loop patterns.MetaProvider for Meta Muse models, ApprovalReviewer for automated pre-approval of low-risk tool calls, and MCP Apps tool results forwarding to AG-UI and Vercel AI frontends.MetaProvider is the first day-one framework integration for Muse Spark Open — if you use pydantic-ai and want to evaluate the Muse 816B model against your Claude or Gemini baseline, upgrade to v2.57.0 and swap the provider. The ApprovalReviewer is an interesting agent pattern: configure a "reviewer model" that runs before human approval; tool calls the reviewer model marks low-risk proceed automatically, while high-risk or uncertain calls escalate to a human. This reduces human review fatigue in high-volume agent pipelines. MCP Apps tool results forwarding means pydantic-ai agents using MCP tools can now surface results directly in AG-UI or Vercel AI chat frontends without custom plumbing.--image-pull-secret for deploying through a private registry. CLI 0.4.36 (Oct 16) extends deployment aliases: langgraph deploy now accepts both serverless and dedicated deployment aliases (not just one type), and pushes through the active Docker context rather than requiring a Docker daemon running locally.langgraph deploy work through the active Docker context and accepts both serverless and dedicated deployment aliases — useful for CI/CD pipelines that build in remote Docker contexts.DOCKER_HOST), langgraph deploy now works correctly in that setup rather than requiring a local daemon. The deployment alias change means teams that have both serverless and dedicated LangGraph deployments no longer need to specify separate alias formats per deployment type. The Oct 14 --image-pull-secret flag is also worth noting if your LangGraph deployment pulls from a private registry.langgraph deploy in CI/CD pipelines; you use remote Docker contexts; you manage both serverless and dedicated LangGraph deployments.EffortQuick (CLI upgrade, no graph code changes).ANTHROPIC_BASE_URL or equivalent at http://localhost:8787/bili/<real-api-url>, and the proxy compresses conversation context by injecting four model-usable tools (compress, decompress, search_context, acp_status). The model decides when to compress. The design maintains prompt-cache prefix alignment. Try it low-stakes first — it's early software with known rough edges.No confirmed benchmark changes in the Oct 15–16 window. arena.ai leaderboard data unavailable (DNS failure). Section omitted.
between_tools to thinking types in model capabilities — a formal declaration that a model supports interleaved thinking (reasoning between individual tool calls, not just at the start). This is part of a broader pattern: the SDK now exposes model capabilities as queryable properties rather than requiring developers to maintain external capability matrices. The speed parameter addition is the same pattern applied to fast mode. Neither requires the developer to know model-specific rules ahead of time.between_tools thinking," "I can run at fast speed" — lets agents adapt their behavior at runtime without code changes when new models ship. This could significantly reduce "which model supports what" maintenance burden in agent frameworks.simonwillison.net DNS failure. hnrss.org DNS failure. No qualifying HN threads with score >200 surfaced in the October 15–16 window. Nothing cleared the quality bar this period.
managed-agents-2026-04-01 header). No GA date announced.Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.