Cheap speed, a parse router, and AI that lives on the laptop

Anthropic released Claude Haiku 5.5 on October 7 as a small model for high-volume, cost-sensitive work: summaries, classification, database queries, compaction, live customer support and browser use. The company says it also works as a subagent beside Opus 5.5 and Sonnet 5.5 on coding jobs. Developers call it as `claude-haiku-5-5` on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.

List prices for prompts up to 100,000 tokens are $0.10 per million input tokens and $0.50 per million output tokens, with cache reads at $0.01. Prompts over 100,000 tokens cost $0.50 / $2.50. Anthropic says that cheaper band covered about 90% of Haiku 4.5 traffic, that Haiku 5.5 is 90% cheaper than 4.5 on those requests and 50% cheaper above 100,000 tokens, and that after tokenizer changes the average cost to run work is about 75% lower. Those figures are the company’s, not an independent audit.

Anthropic’s published scores show a large jump over Haiku 4.5: 72.4% versus 15.7% on OSWorld 2.1 (offline subset), 39.2% versus 0% on Terminal-Bench 4.0, and 45.9% versus 10.2% on Humanity’s Last Exam without tools. On those tables Haiku 5.5 also beats GPT-6 Luna on several agent and knowledge-work tests, while remaining behind Sonnet 5.5 on the hardest coding benches. Customers quoted on the launch page reported faster turns: Asana said more than 30% lower task latency and up to 2.5× faster inference per agent turn; HubSpot said 92.8% on an internal CRM suite.

The rest of the Claude stack moved with it. Cache reads on Sonnet 5.5 were halved to $0.10 per million tokens, which Anthropic says cuts most agentic Sonnet work by about 20%. Max and Team subscribers will get monthly API credits ($100 for Max 5x, $200 for Max 20x, up to $500 pooled on Team). Python and TypeScript SDKs are adding computer-use and browser-use in beta, which Anthropic says Haiku 5.5 is well suited to.

Caveats matter. Benchmarks and customer quotes come from Anthropic. Cybersecurity safeguards are tighter than Haiku 4.5 but looser than Sonnet 5.5: defensive work is allowed, penetration testing is blocked unless an organisation is in Anthropic’s verification programmes. Sonnet and Opus remain the models Anthropic recommends for complex agentic coding. The commercial opening is for the long tail of cheap, repetitive agent steps that used to be too expensive to run at volume.

Why it mattersA fast small model at roughly a tenth of last generation’s typical-prompt price makes it realistic to put a subagent on every lookup, classification and support turn instead of reserving frontier models for everything.

LlamaIndex launched OpenDocRouter on October 7 as a hosted document-to-Markdown API. A Hugging Face search for “ocr” already returns thousands of models, and labs ship document-capable models almost monthly. OpenDocRouter’s pitch is that teams should not rebuild prompts, rate limits, hosting and benchmarks each time. Call `POST /v1/parse` and swap the model.

At launch the lineup is five frontier models (Claude Opus 5.5, Gemini 3 Flash, Gemini 3.8 Flash, GPT-5.6 Terra, GPT-6 Luna) and five open-source ones (Infinity-Parser2-Flash, MinerU2.5-Pro, TeleOCR, dots.mocr, PaddleOCR-VL-1.6). LlamaIndex scores them on ParseBench across tables, charts, faithfulness, formatting and grounding. Claude Opus 5.5 leads overall at 84.20, including 93.53 on tables, at a listed $48.82 per 1,000 typical pages. GPT-6 Luna is 71.34 overall at $0.80 per 1,000 pages; MinerU2.5-Pro is 70.05 at $0.86. Those page costs are LlamaIndex’s token-use estimates on ParseBench documents as of October 6, not a promise for every invoice.

Mechanically, each page is its own model call with its own Markdown, status and charge. Failed pages retry; callers can skip pages they already have. Files can be a public HTTPS URL or an upload, capped at 50 MB or 500 pages (inline base64 about 3 MB). Up to 50 pages run synchronously; larger jobs go async. Optional `layout: true` adds bounding boxes in reading order via LlamaIndex’s grounding engine at $0.20 per million tokens on pages whose layout succeeds. Nothing is retained unless caching is on; cached results last 24 hours encrypted.

Billing is prepaid credits from $25 (5% top-up fee). Frontier models are billed at provider token prices with no markup, LlamaIndex says. Failed, cached and blank pages are free. Account limits include 10 concurrent requests and 300 parse calls per minute. LlamaIndex distinguishes this from LlamaParse: OpenDocRouter is the swap-and-compare layer; LlamaParse keeps hand-tuned tiers, enterprise controls and extra APIs such as schema extraction.

The opportunity is not another OCR wrapper. It is routing: send invoices to a cheap open parser, send 100-page financials with nested tables to Opus, and prove the quality gap with a shared benchmark. Teams that still run a single parser for every document are leaving both money and accuracy on the table.

Why it mattersDocument AI is no longer one model. A router with public quality and cost numbers lets operators treat parsing like they already treat chat models: cheapest model that clears the bar.

At a Windows and Surface event in San Francisco on October 7, Nvidia and Microsoft said they are co-engineering PCs for AI agents that run on the machine. RTX Spark laptop preorders opened the same day, with availability from October 16; compact desktops are due in November. Nvidia lists a Blackwell RTX GPU with up to 6,144 cores and a Grace CPU with up to 20 cores, linked at 600 GB/s, with up to 128GB of unified memory and 1 petaflop of FP4 AI performance.

Microsoft’s Pavan Davuluri said Surface Laptop Ultra is built around RTX Spark so models that do not fit a typical laptop can run locally. Reporting based on Tom’s Hardware and Thurrott put Surface Laptop Ultra from $2,599, the Surface RTX Spark Dev Box from $5,999, and HP’s OmniBook Ultra 16 from $3,199. Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte are in the first wave. Microsoft’s own comparison versus an M5 Pro MacBook Pro (up to 2.1× on first text, 4.3× on pictures, 6.2× on clips) is a vendor claim, not an independent test.

The software half is Microsoft Execution Containers (MXC), now generally available. MXC is OS-level containment so agents can run in the background under policy: which files and network destinations they may use. GitHub Copilot, OpenAI Codex, Replit, LM Studio and others already support it; Anthropic Claude Code and several more are listed as coming. Nvidia is integrating OpenShell with MXC for extra policy, credentials and enterprise audit logs.

Nvidia also previewed DGX Station for Windows: a deskside box on the GB300 Grace Blackwell Ultra Desktop Superchip, with 748GB of coherent memory and up to 20 petaFLOPS of FP4 compute, aimed at trillion-parameter-scale local work. Until now DGX Station was a Linux machine, which forced many Windows-standard enterprises to keep two environments.

This is not a cheap PC refresh. It is a bet that developers, creators and some enterprises will pay a premium to keep weights, files and agents on the desk, under OS policy, instead of sending everything to a cloud model. Supply, thermals, battery life and real agent sandboxing will decide whether that bet holds after the keynote.

Why it mattersLocal agents only become a default product if the PC can hold a large model and the OS can jail it. RTX Spark plus MXC is the first mass-market attempt to sell that combination as a Windows SKU.