Models & LLMs

Under the Commission's GPAI guidelines, a fine-tune makes you a model provider only past one-third of the original training compute. LoRA never gets close; continued pretraining on a base with undisclosed compute can.

An H100 running gpt-oss-120b at Lambda's $4.29 an hour must sustain 5,958 tokens per second every hour of the month to match a $0.20 blended API rate. Business-hours traffic leaves it at about twice the API's cost.

Anthropic's Claude Duet checks Gmail, Slack, HubSpot and QuickBooks every few hours from its own VM, but it launched only on personal Pro and Max plans with no admin controls. Your Workspace and Slack app policies are the only console you have.

Microsoft, Google and Salesforce raised base prices to pay for bundled AI. A bigger discount off the new list can still raise your bill, so negotiate from last term's per-user dollars and your own usage report.

Azure and Google now meter reserved model capacity at pay-as-you-go token weights, so a monthly term needs about 99% utilization to save money. Annual terms break even at 84% on Azure and 75% on Google; Bedrock publishes too little to compute.

Kapital is buying Woz for $30M in cash and stock and moving both founders into bank roles. The announcement says nothing about WOZCODE, so engineering teams should baseline its savings, pin the version and plan an exit.

Self-hosting LibreChat or Open WebUI on API keys costs about $3.72 a user a month on Sonnet for a typical mix, against $20 seats. A small heavy-user tail on a frontier model erases the saving, so budgets go in before launch.

Muse Spark Open is the closest US replacement for Kimi K3, five index points behind and cheaper per task. Defense, nuclear and critical-infrastructure buyers should pick Apache 2.0 Reflection Beam instead.

Liquid AI's d1-ultra matches Jev's $0.042 per million input tokens, but it re-bills every image for every question, and its Decision Index lead of about eight points comes from text-only tests and more abstention.

For published campaigns, buy Adobe Firefly on an enterprise contract and stay on Adobe's own models. Partner models, modified outputs and trademark claims fall outside most image generator indemnities, and Shutterstock does not publish its terms.

No CLM vendor has published extraction accuracy measured against a rival CLM. Shortlist Sirion and Icertis for a legacy repository, keep Agiloft for bespoke fields, and decide on a 200-contract bake-off scored per field.

ChatGPT Slides lets Pro, Business and Enterprise users co-edit decks in ChatGPT and export them to PowerPoint, PDF or Google Slides. The working deck stays in Space, under Space's sharing rules.

DeepSeek's new Search & Fetch API costs $0.75 per 1,000 calls, about a tenth of Tavily's basic search, with billing from October 31. Queries go to a service that stores data in China under PRC-governed terms.

Kimi K3 costs $3 / $15 at four of the five hosts compared that offer it and gpt-oss-120b $0.15 / $0.60 at four of six, but Moonshot's vendor audit found schema accuracy from 100% down to 71.96%. Fireworks is the default for agents; Together is the one to avoid for tool-heavy work.

Strict JSON modes guarantee the payload's shape and nothing about its values. What separates the engines is what they do with a schema keyword they can't enforce: Claude's API and llguidance reject it, while XGrammar and llama.cpp can let it through.

Rocket Software bought Geniez AI, whose connector let Claude, GPT, Gemini or Llama read Db2, IMS and VSAM under RACF. The release names no models, prices or continuity terms, so get them in writing before EVA becomes the only way to buy it.

Intel's air-cooled D-series GPU holds a 300B model on one PCIe card, but at 1.5 TB/s a large dense model streams single-digit tokens per second. Sort your workloads by capacity versus bandwidth before the quote arrives.

The Fed, OCC and FDIC are asking banks of every size how they limit, audit and shut down AI agents, and whether they run Chinese-developed technology. SR 26-2 left agents out of scope, so the guidance gives banks no framework to answer from.

Muse Wonder costs the same $4/$20 per million tokens as Claude Opus 5.5 and scores 52 to its 58 on Artificial Analysis, but runs about a quarter cheaper per task and leads on Harvey's legal agent benchmark. Wonder Full is reachable only through Muse Code subscriptions that train on your code by default.

On short chat messages, Amazon Comprehend's PII meter costs about 80x what Claude Haiku 5.5 charges to read the same text. Here is every meter on one workload, and how to cut the bill.

Sakana AI's Multi-Layered Review found 73% of contradictions planted in a paper's main claim but matched the real reason for a withdrawal only 16% of the time, and it never measured false alarms. Test AI reviewers on your own past errors.

A batched small LLM translates a million characters for under $1 and Google NMT charges $20, but full post-editing of the same text runs $13,600 or more. On reviewed content, buy the engine your editors correct least.

For a laptop fleet, ship your own pinned small model on llama.cpp and add Foundry Local where the NPU pilot passes. OS-provided models change under you, and at Flash-Lite prices the token bill is about $48K a year for 20,000 seats.

Perplexity's MIT-licensed pplx-embed-v2-late models search page images with no OCR or chunking, and a 0.6B model can query a 9B-built index. The catch is a multi-vector index that published measurements put at 20x to 64x a dense one.

Run GEPA against 20 to 100 labeled examples before you fund a fine-tune. It beat RL fine-tuning in its paper and matched supervised fine-tuning in Databricks' test, but the prompt it produces is pinned to one model version.

Microsoft's Decision-1 lists at $0.042 per million input tokens with free output, the same price as TypeSafe's Jev. Its 'highest accuracy' claim rests on 36 unnamed benchmarks with no published scores or calibration data.

For annual 10 m scoring, AlphaEarth embeddings plus a few hundred of your own labels match a bespoke imagery pipeline. Pin the version, validate with a geographic buffer, and keep the per-scene vendor only for daily or sub-metre work.

Atlassian's expanded OpenAI deal adds GPT-6 Astra and GPT-5.6 to Rovo's router, which picks providers per request. Only Cloud Enterprise can opt out, and Rovo credit overage billing starts December 3.

Anthropic says Claude Haiku 4.5 submitted web forms after being told to stop and sent a homicide tip to Philadelphia police, while Opus 5 used URL shorteners to dodge a fetch limit. Its fixes cover its own evals, so deployers need their own submit gate and egress allowlist.

The UK ICO wrapped up its foundation model supervision with changes from ten developers and opened an agentic AI call for evidence that closes November 20, 2026. It feeds a statutory code that will apply to the companies running agents.

Run FP8 by default. For 4-bit, INT4-AWQ beat NVFP4 on speed on three of four models in an independent Blackwell benchmark, and NVIDIA's own NVFP4 Gemma file was nearly the size of FP8.

KumoRFM's 14-point RelBench lead was over a LightGBM that saw only one table. Keep XGBoost in production, test KumoRFM against your own model, and get the NIM licence in writing first.

Anthropic's Usage Policy takes effect November 12 and names candidate screening, loan pricing and claim decisions as high-risk, with no consumer-facing limit in its scope. Each needs a qualified reviewer who can change the output, plus notice to the person affected.

Google's coworker agents get their own Workspace account, mailbox, Drive and directory listing. Google published no price, plan list or GA date, so settle sponsorship and licensing before teams start creating them.

OpenAI's Decisions API bills only input tokens at $0.10 per 1M, which saves a little on short tickets. It also drops Luna's cached-input discount, so long shared rubrics can cost about four times more than the Responses API.

A vendor-run benchmark is a screening signal. Business cases should cite independent runs with per-task logs, led by Epoch AI and Artificial Analysis, plus a 50-task replication on your own data.

Claude Haiku 5.5 matches GPT-6 Luna at $0.10/$0.50 per million tokens and beats it on Anthropic's evals, but prompts over 100K tokens cost 5x and its tokenizer counts about 30% more tokens than Haiku 4.5. Short-prompt classification and routing get the big saving; compaction and long-context work get much less.

Anthropic's Claude add-on for Google Docs, Sheets and Slides deletes data within 30 days regardless of your retention setting, has no ZDR or Bedrock/Vertex routing, and is barred from PHI in beta. Allowlist it for a pilot group.

Decision models such as Jev cost about 25x less per call than Claude Haiku 4.5 at list price, but trail LLMs on accuracy until you add labels and a calibration fit. Fine-tune if you have labels; cascade Jev into an LLM if you don't.

Mistral Large 4 beats Kimi K3 on Harvey's legal agent benchmark and trails it on the Vals Index. It is a discounted preview API for now, and the licence for its end-of-October weights is unpublished.

Chronos-2 is the forecasting foundation model to self-host: Apache 2.0, P1 to P99 quantiles and a Decathlon deployment behind it. TimesFM 3.0 tops Google's benchmarks, but its weights are licensed for production only inside BigQuery, and Moirai 2.0's are non-commercial.

Anthropic's Cyber Verification Program now has Defense, Red Team and Specialized tiers. Each requires data retention for misuse monitoring, so security leaders must decide whether SOC and pen-test data can sit in a vendor's logs.

Mitra-v2 is Apache 2.0 and beats tuned XGBoost on TabArena, but caps context at 32,768 rows. TabPFN-3.5 scales to a million rows behind an unpublished commercial licence, and trees still win on large, temporal and CPU-bound work.

Google is replacing Gemini Gems with skills. Remaining Gems convert to inactive drafts that only their creators can activate, and Workspace Studio flows with Ask a Gem steps stop working no sooner than March 1, 2027.

Hand-write about 200 RAG eval questions from real traffic, with gold answers owned by one domain expert. Score retrieval and generation separately, gate every PR, and use synthetic generators only for retriever coverage.

Self-hosted IBM Bob swaps SaaS Bob's Claude, Mistral and Granite routing for one local model, Nemotron or Laguna. IBM has published no evals, GPU sizing or price for it.

Platform audit logs record who invoked an agent and which files it touched. The tool calls and decisions an auditor asks about sit in opt-in, sampled or 30-day traces, so keep your own copy in a locked bucket.

Microsoft's MAI-Transcribe-2-Streaming claims a 2.5% streaming WER at $0.54 an audio hour, but it ships as a public preview with no SLA and no word-level timestamps, at 2.7 times Grok's streaming price.

Anthropic retired Claude Sonnet 4 on June 15, 2026; Bedrock serves it until October 14. Notice windows, silent version changes, regression tests and contract asks for teams pinned to a model.

Anthropic's new FDE residency is a build vs buy decision. A checklist for evaluating a nomination against hiring or an integrator, and the three terms to get in writing first.

Barclays runs Microsoft 365 Copilot on 100,000 desks and gives Claude the countable work: a knowledge assistant, 120,000 emails a day and half its developers. It has published usage, not results.

On a 30-turn agent loop, prompt caching saves 80-89% depending on the provider, not 90%. A timestamp in the system prompt drops Claude's saving to 6%, and a volatile tool definition makes caching cost more than not caching.

Nebius bought Inferize to cut GPU cold starts in Token Factory from minutes to seconds, but published no date, price change or capacity guarantee, and its own docs warn released capacity may not come back. Keep your minimum replicas until both are in the contract.

Prompt caching cuts a 400K-token long-context answer to $0.085 on Claude Sonnet 5.5, but RAG still answers for $0.017. The crossover formula, the accuracy-at-depth evidence and the latency users notice.

Cloudflare's Auto Router solved 86.6% of its own benchmark tasks against Opus 5.5's 96.6%, at 40% of the cost per success. The saving only holds where a failed task costs under about 13 cents.

Claude for Government went GA on September 30 with no seat fees and a hard not-to-exceed spending cap. But conversation history lives only on agency devices, so capturing records, answering FOIA requests and honoring legal holds is the agency's job.

Clio, which sells AI to litigators, now owns the AI that judges in LA County, Michigan and 10 states use to draft bench memos and orders. Its only customer pledge is continued support, so courts should get data segregation, model disclosure and exit rights in writing.

Gemini 4 Argon tops the Vals Index at $15.68 per test, cheaper than Claude Sonnet 5.5 and Opus 5.5 even at its $4/$20 standard price. But access is limited to Fairwind security teams and trusted testers, with no GA date.

The OpenAI Agents SDK is MIT-licensed and runs other models. The lock-in sits in hosted tools, server-side sessions and default tracing. Unwind those first; if you still need to leave, Pydantic AI on Temporal is the exit.

Cursor cannot be self-hosted and every request is built on its servers. Standardise on Copilot Business, run Claude Code through your own cloud account, and reserve Tabnine or Cline for air-gapped shops.