Claude Haiku 5.5 lands at a tenth of Haiku 4.5’s price.
Anthropic ships Claude Haiku 5.5: a 1M-context small model at $0.10 in, $0.50 out
Haiku 5.5 has a 1M-token context window, up to 128K output tokens, and is the first Haiku with the effort setting, so you can trade quality for speed and cost per request. Anthropic pitches it for classification, routing, extraction, summaries and compaction, and as a fast subagent under Opus 5.5 or Sonnet 5.5. Anthropic says it is a big step up from Haiku 4.5 and costs about 75% less to run on average. The benchmarks and customer quotes in the launch post come from Anthropic.
The price drop is real, but this is not a drop-in swap. Adaptive thinking is on by default, so a response can start with thinking blocks. The old fixed thinking budget now returns an error, and so do custom temperature or top-p values and prefilled assistant turns. A new tokenizer counts the same text as roughly 30% more tokens. Computer use moves to a new toolset, Priority Tier isn’t supported, and the model can refuse requests with no server-side fallback. Anthropic’s migration guide lists each change, and Claude Code ships a migration skill.
Anthropic made more pricing changes alongside the launch. Cache reads on Sonnet 5.5 are now half price, which Anthropic says makes most agentic work about 20% cheaper. Max and Team subscribers get monthly API credits starting this week. The Python and TypeScript SDKs add beta classes that run browser-use and computer-use loops against your own automation.
What to do: re-count tokens on a sample of real prompts before you trust the 75% figure, strip sampling parameters and prefills, and make your client pick content blocks by type. Haiku 4.5 has no retirement date yet, so you can migrate on your own schedule.
Sources: Anthropic — Introducing Claude Haiku 5.5 · Claude release notes (Oct 7) · Migration guide · AWS What’s New
| Model / package | Org | Notes |
|---|---|---|
| Claude Haiku 5.5 Oct 7 · GA |
Anthropic | Cloud 1M context · effort · $0.10 / $0.50 per 1M (≤100K) |
| Mellum2.1 Oct 8 |
JetBrains | Open 12B MoE / 2.5B active · Apache 2.0 · agentic coding |
| d1-3B / d1-omni-600M Oct 7 |
Liquid AI | On-device decision models · LFM 1.0 license · omni experimental |
| LiteRT 2.3.0 Oct 7 |
Google AI Edge | On-device Metal fused attention · residency · Swift API |
| Ollama 0.40.1 Oct 7 · patch |
Ollama | GA Windows fixes · MLX residency patch now upstream |
| Nemotron IMO / IOI specialists Oct 7 |
NVIDIA | Open SFT + RL checkpoints · datasets · proof benchmark |
| MAI Code 1.1 Flash (local) Oct 7 · announced |
Microsoft | Watch 3-bit local build · 256K context · Windows |
| mlx-vlm / MLX main Oct 6–8 · unreleased |
Blaizzy / ml-explore | Watch Clef decision support · Metal residency + quant matmul fixes |
| llama.cpp tips Oct 8 · not GA |
ggml | Watch reranker classifier activations · GA still v0.6.0 |
| Item | Beat | Status |
|---|---|---|
| vLLM GHSA-h3rc High 8.1 · Oct 6 · correction |
Request-chosen processor code revision → RCE with remote code on · fixed in 0.31.0 | Active |
| Managed Agents web_fetch Oct 7 |
Only URLs already in the session · host allowlist applies to web tools | Active |
| Microsoft Execution Containers Oct 7 · GA |
OS-enforced agent policy · Windows, macOS, Linux backends | Active |
| Haiku 5.5 safeguards Oct 7 |
Cyber stricter than Haiku 4.5 · blocks pentesting · CVP for more | Active |
| Gemini Agent Gateway Oct 8 · announced |
Org-wide policy on all agent traffic · per-agent identity | Watch |
| Claude Compliance API Oct 8 · beta |
Chat endpoints now cover the unified Claude experience | Active |
| Claude Models API server tools Oct 6 · correction |
Reports web search / code execution support per model | Active |
| Sonnet 4.5 retirement Claude API · Nov 30 |
Migrate to Sonnet 5.5 · no Haiku 4.5 date yet | Watch |
Critical · vLLM remote code · Haiku client breakage
If any vLLM server runs with --trust-remote-code, confirm it is on 0.31.0 or later and keep /tokenize off the public internet. Before anyone switches a config to Haiku 5.5, remove temperature and prefill from those calls, select response blocks by type, and handle refusals.
Watch · RTX Spark ship date · Gemini agent details · Mistral weights
RTX Spark laptops ship Oct 16, and compact desktops follow in November. Google hasn’t given the Gemini agent’s availability or price. Mellum2.1’s GGUF and vLLM speculative-decoding head are “coming soon.” Mistral Large 4 weights are still due by the end of October. NVIDIA’s GTC Berlin keynote is Oct 21. Copilot’s local-model features reach Copilot+ PCs “in the coming months.”
Opportunity · cheaper caching · API credits · permission-aware apps
Sonnet 5.5 cache reads are half price. Max plans get $100 or $200 a month in API credits and Team plans up to $500. Databricks Apps can now act as the signed-in user so Unity Catalog permissions apply (GA). Google’s Developer Knowledge API gives agents current Google docs. Cloudflare wrote up a SOC harness that triages alerts with the open Clef decision model.
Dry-run your Haiku 5.5 migration on 50 real prompts
Take 50 prompts from your logs and count tokens with the model set to Haiku 4.5 and then to Haiku 5.5. Multiply by each model’s prices to get your actual saving. Then send them to Haiku 5.5 with sampling parameters and prefills removed, and check what share come back as refusals or start with a thinking block.
Race Liquid’s d1-3B against Clef on your Mac
Load d1-3B in Transformers on the MPS device (Liquid’s post has sample code) and ask the same three yes/no questions about 50 support tickets that you put to Clef on Ollama’s MLX engine. Compare accuracy and per-question latency. Because d1 runs remote code, run it in a throwaway virtual environment or a sandbox.
Audit your inference servers for remote-code flags
List every vLLM, SGLang or Transformers server you run and note which start with remote code trusted. For each one, record the version and check whether /tokenize is reachable without a key. Upgrade vLLM to 0.31.0 or later, pin model revisions, and turn the flag off wherever the model doesn’t need it.
Issue 34 · Thursday, October 8, 2026 · Not investment advice · Sources cited inline · Window: after Issue 33 (Wed Oct 7 morning ET) through Thu Oct 8 morning ET; Commentary Oct 6–8 · Haiku 5.5 was a watch item in earlier issues; launch is the new delta · Correction: vLLM GHSA-h3rc (Oct 6, High) was missed in Issues 32–33 · Correction: Claude release notes have an Oct 6 entry (Models API server tools) that Issue 33 didn’t list · llama.cpp tips after v0.6.0 are not GA · mlx-vlm and MLX main are unreleased · Mistral Large 4 weights not yet released · Listen omitted (episode not published)
Tuesday through Thursday morning (Oct 6–8 ET). ASTGL / Fraser takes, sourced, not recap theater.
The small-model price war just hit a dime
Haiku 5.5’s list price for short prompts, $0.10 per million input tokens and $0.50 output, matches what OpenAI charges for GPT-6 Luna. Two labs now sell their cheap tier at the same headline price, and Anthropic is pitching Haiku as the subagent under a bigger model. That changes the local-versus-cloud math for the boring jobs: summarizing, routing, tagging, compacting. A local model no longer wins those on cost alone. It wins on privacy, offline use, and latency you control. Don’t compare price sheets, though. The two models use different tokenizers, and Anthropic’s own docs say Haiku 5.5 counts the same text as about 30% more tokens than Haiku 4.5. The “75% cheaper” figure is Anthropic-reported. Run a day of real traffic through each and compare the invoices.
Sources: Anthropic — Haiku 5.5 · What’s new in Haiku 5.5 · OpenAI API changelog (GPT-6 Luna pricing)
The cheap model is the one with the breaking changes
What I notice most in this launch is how much old client code breaks. No custom temperature, no prefill, no fixed thinking budget. Thinking is on by default and can be the first block in a response, and refusals come back with no fallback. Haiku is the model people wire into pipelines and then forget about, so it’s the one most likely to be running on years-old request code. If you have a classifier that reads
content[0]and setstemperature: 0, it will fail on day one. The fix is small, but it has to be made everywhere. Search your repos for the Haiku model ID before somebody bumps it in a config file.Sources: Haiku 5.5 migration guide · Claude release notes (Oct 7)
Windows wants to be your local AI box
Wednesday’s Microsoft and NVIDIA event was the clearest pitch yet for a Windows PC as a local-model workstation. RTX Spark laptops go on preorder with up to 128GB of unified memory and ship October 16. Microsoft is bringing a 3-bit build of its MAI Code 1.1 Flash coding model to run locally, and it says Copilot will route work between local and cloud models. A DGX Station for Windows was previewed with 748GB of memory. If you run models on a Mac, that is the comparison the hardware makers are now inviting. Unified memory is no longer something only Apple sells, and CUDA comes along with it. NVIDIA’s line that the local models match “many cloud models” is an NVIDIA claim. My take: memory size per dollar and the runtime you already use (MLX or CUDA) matter more than the petaflop number. Wait for independent tests on real 70B-class workloads before you switch platforms.
Sources: NVIDIA (Oct 7) · Windows Experience Blog (Oct 7)
Agents finally get walls they can’t argue with
Three platforms shipped the same idea in two days: the agent doesn’t decide what it can touch. Microsoft made Execution Containers generally available, with policy that sits outside the agent and a process sandbox that uses Seatbelt on macOS, so it isn’t only for Windows. Anthropic’s Managed Agents now only fetch URLs that already appeared in the session, which closes a quiet exfiltration path where a tool’s output contains a URL and the agent fetches it with your data in it. Google’s new Gemini agent sends all agent traffic through a gateway that enforces org policy. That’s the right direction. Prompts don’t enforce anything. If your local coding agent runs with your full user account today, this is the week to put it in a sandbox, even a plain OS one.
Sources: Windows Developer Blog — MXC · Claude release notes (Oct 7) · Google Cloud — Gemini at Work
We missed the worst vLLM advisory of the week
A correction first. Issues 32 and 33 covered four vLLM advisories published Oct 6. There was a fifth, and it was the most serious. On servers started with
--trust-remote-codeand serving a model whose processor ships its own code, a request could choose which revision of that code gets imported. That allowed remote code execution, and/tokenizesits outside--api-key. It is rated High (8.1) and fixed in 0.31.0, the same upgrade we led with on Monday. The default configuration isn’t affected. Remote-code models keep shipping, though: Liquid’s new d1 models needtrust_remote_codetoo. If you serve anything with that flag on, check your vLLM version today and put/tokenizebehind a proxy.Sources: vLLM GHSA-h3rc-6mm3-gc2m · Liquid AI — open d1
Apple GPUs keep getting better kernels from other people
Google’s LiteRT 2.3 added FlashAttention-style prefill and decode kernels on Metal and uses Apple’s residency API to keep model weights and KV cache in RAM. The same evening Ollama dropped a Metal residency patch it had been carrying “now that it is upstream,” and earlier this week MLX main gained a setting for how often residency refreshes. Google reports about 38% faster decode on a small Qwen and nearly 3× faster 32K prefill on an M4 Pro (Google-reported). The pattern is the story: a lot of the Mac’s local-AI speed now comes from memory management and attention kernels written outside Apple. If you build on-device apps, LiteRT’s new Swift API is worth an afternoon. If you run Ollama, update to 0.40.1 and re-check memory pressure on long contexts.
Sources: LiteRT v2.3.0 · Ollama v0.40.1 · MLX main