BuildRiot
Sign in

AISource: Anthropic

Anthropic moves Claude Opus 5.5 fast mode out of research preview

Anthropic removed access requests and the beta-header requirement for Claude Opus 5.5 fast mode. Default fast mode rate limits cover Start, Build and Scale, and Azure-hosted Microsoft Foundry deployments are supported.

By News DeskPublished here (UTC)1 min read
Anthropic logo
Cover: BuildRiot

Anthropic announced on October 15, 2026, that Claude Opus 5.5 fast mode had exited research preview, removing access requests and the beta-header requirement. Requests that retain the fast-mode-2026-02-01 header continue to work. Sources: Anthropic

Default fast mode rate limits apply to Start, Build and Scale; Evaluation organizations must move to one of those tiers to use it. Sources: Anthropic

Developers enable fast mode on the Claude API by selecting claude-opus-5-5 and setting speed to "fast". Anthropic claims up to 2.5 times the output tokens per second with the same model weights and capabilities. That claim concerns generation throughput, not time to first token. Sources: Anthropic

Pricing is $8 per million input tokens and $40 per million output tokens, including requests exceeding 200,000 input tokens. Caching and residency modifiers apply. Fast mode has separate limits, and applications must handle fallback to standard speed. Cached prefixes are not shared across speeds. Sources: Anthropic

Opus 5.5 fast mode is also available through Claude in Microsoft Foundry for Azure-hosted deployments. Anthropic-hosted Foundry deployments are excluded. Opus 5 and Opus 4.8 fast mode remain in research preview. Both still require the beta header. Sources: Anthropic

Why it matters

Developers on Start, Build and Scale can use Opus 5.5 fast mode without preview access approval. Premium pricing, separate capacity limits and application-managed fallback remain tradeoffs.

Attribution

Sources

  1. 1AnthropicClaude Platform release notesSource published (UTC)
  2. 2AnthropicFast mode
  3. 3SessionWatcherClaude Opus 5.5 fast mode leaves research preview with default API rate limitsSource published (UTC)

Read the original source for full details and context.

Report a problem

Links, dates or what you expected. Do not include passwords or private data.

More on AI

All AI stories

Review · AI

Elastic’s Linux profiling preview: Setup, requirements and limits

Elastic says its OpenTelemetry Profiling integration collects continuous CPU profiles on Linux through eBPF. It requires no application instrumentation or restarts, according to Elastic. Version 0.0.1 is a technical preview. OpenTelemetry Profiles remains Alpha. Elastic’s October 14, 2026, walkthrough performs the one-time Universal Profiling storage initialization as superuser. It then configures a Fleet policy with samples_per_second=20 and enrolls the Linux agent. To verify collection, Elastic directs users to Infrastructure > Universal Profiling > Flamegraphs. Users select the traffic window and filter process.executable.name: "java"; agent health alone is insufficient. Elastic reports 30,150 PetClinic requests in four minutes at slightly over half a CPU core. The walkthrough does not report a profiling-overhead measurement.

Elastic

Review · AI

Nemotron Bolt claims a 24GB footprint; its deployment recipes target larger GPUs

NVIDIA positions Nemotron Bolt for local agentic work, including coding, with a stated 24GB memory footprint. Its published GGUF deployment recipes target discrete GPUs with at least 32GB or 48GB of memory, depending on configuration. NVIDIA announced availability on October 15, 2026. The quantized model card lists 74 billion total parameters and 9 billion active parameters. It supports multimodal inputs and text output. It specifies mixed quantization: IQ2-XXS experts, NVFP4 Mamba layers and Q8_0 embeddings and output. The setup instructions use llama.cpp with a separate LiLiCorr drafter for speculative decoding.

NVIDIA