ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 23, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Anthropic's new Claude API version returns an error on any request that sets temperature, top_p or top_k, and the upgraded SDKs send that version by default. Today is about what an SDK upgrade can break before anyone tests it. Amazon also saw a PHP exploit attempt four days before disclosure, and OpenAI priced a decision model at $0.10 per million tokens.

By the numbers

  • 3 changes Anthropic's guide: temperature settings, document blocks in assistant messages, and citations in user text blocks return a 400 error.
  • 4 days Amazon says it saw PHP2Shell tried on 17 October; the flaws went public 21 October. Its own honeypot data.
  • $0.10 per 1M tokens OpenAI says output is not billed. Two public indices rank its model eighth and fourteenth, per TechCrunch.
  • 1,000 of 30,000 Uber's account: about 3% of staff piloted ChatGPT Work first; no cost or time-saved figures were published.
The upgrade that fails

Anthropic's API version 2026-10-21 rejects temperature settings, and its new SDKs send it by default

Anthropic released Claude API version 2026-10-21 on 21 October 2026, with new major versions of eight client libraries and its command-line tool. Anthropic says the SDKs now send the new version by default: Python 2.0.0, TypeScript 1.0.0, Go 2.0.0 and five others. Its migration guide lists three Messages API changes that return a 400 error, an outright rejection, instead of a warning.

Why it matters: Sending temperature, top_p or top_k now fails even at the default value. Without them a request samples at a temperature of 1.0. Anthropic says models that only accept defaults, such as Claude 4.7 and later, behave as before. On models that accept other values, such as Claude Sonnet 4.6, a request that lowered temperature for steadier output loses that effect, and nothing replaces it. A support agent tuned for consistent answers could start varying its wording without any error you would see.

The other two failures are narrower: a document block placed directly in an assistant message, and a user text block with a non-empty citations array. The SDKs also send your API key as an Authorization: Bearer header instead of X-Api-Key. Anthropic says the API accepts either, but a proxy, gateway or log filter that reads X-Api-Key will stop seeing the key.

The catch: These are Anthropic's own pages, so they describe the intended path. The guide says you can stay on the old version, 2023-06-01, but gives no date for ending it, and I found none. On Amazon Bedrock, the guide says the new version covers only Claude 4.7 and later, so a mixed fleet needs two settings.

What to do: Before any SDK bump reaches production, search your code and gateway configuration for temperature, top_p, top_k and X-Api-Key. Replay 50 recent requests from each application against a staging key on the new version, and compare the outputs for the ones that lowered temperature. Pin the old version for those until the comparison passes.

Our coverage →Anthropic: Migrating to API version 2026-10-21 ↗Anthropic: Claude Platform release notes ↗

Four days early

Amazon says it saw PHP2Shell exploited on 17 October, four days before the flaws were disclosed

Amazon Threat Intelligence said on 21 October 2026 that a sensor in its MadPot honeypot network logged an attempt to exploit PHP2Shell on 17 October. The two PHP flaws, CVE-2026-109618 and CVE-2026-109619, became public four days later, when the embargo lifted at 14:00 UTC. Amazon says the first proof-of-concept it found that day appeared around 02:00 UTC, and that PHP2Shell is now broadly exploited.

The catch: This is one vendor's telemetry from decoy systems. Amazon does not say whether any customer was compromised or how the 17 October actor learned of the bug. Its claim that AI-assisted analysis is shortening the time from patch to exploit is its own assessment; the post measures nothing.

What to do: Ask your platform team which PHP version runs your internal tools, admin panels and AI front ends, including ones a vendor hosts. Patch to the fixed versions, search logs for POST requests containing PHP_SESSION_UPLOAD_PROGRESS, and look back to 17 October, not 21 October. Set a patch target for critical internet-facing flaws in hours, and check whether your policy still says days.

Our coverage →AWS Security Blog: Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot ↗

Fast, not yet best

OpenAI makes its Decisions API generally available at $0.10 per million input tokens

OpenAI said on 22 October 2026 that its Decisions API is generally available, running on gpt-6-luna. It returns typed answers instead of text: a probability, a choice from your list, or a score. A request can carry up to 400 questions. OpenAI's guide says it is about 10x faster than its Responses API and charges $0.10 per million uncached input tokens, with no charge for output.

The catch: TechCrunch reports that two public decision-model indices, the AIM System One Index and the Artificial Analysis Decision Index, place OpenAI's model eighth and fourteenth. OpenAI's page publishes no accuracy figures, and the latency drop, from 1 to 3 seconds to 100 to 180 milliseconds, is OpenAI's own account.

What to do: List the routing, triage and tagging steps in your agents that call a full model only to get a label. For each, build 200 to 500 labelled real cases and compare decision models on accuracy, not speed. Keep a human path for low-confidence answers; OpenAI leaves the threshold to you.

Our coverage →OpenAI: Decisions guide (question types, pricing, availability) ↗TechCrunch: OpenAI makes its Jev competitor widely available as decision model competition heats up ↗

What else

  1. 01Uber told TechCrunch it is giving all 30,000 employees ChatGPT Work after a 1,000-person pilot, about 80% of it non-engineers; it disclosed no spend or success metric. Our coverage →
  2. 02Rescale says its Bedrock triage agent for nightly test failures rose from 44% to 68% agreement with human triage and cut deploy-blocking time 20%; the figures are from AWS's vendor-published write-up. Our coverage →
  3. 03A Hugging Face community post reports that sorting similar prompts together before an offline vLLM job raised mixture-of-experts throughput 12% to 17% on one GPU. Our coverage →

Stat of the day

25% of pull requests

The Register reports a Linux Foundation survey in which respondents estimated generative AI cut code-authoring time 52% on average, while about 25% of incoming pull requests were low quality, 31% on large projects. Only 35% of projects have an AI policy. These are estimates from respondents, not measurements. The cost of faster writing lands on reviewers, so set disclosure rules and review capacity before a wider coding-assistant rollout. The Register: AI cuts coding time but is also filling open source's inbox with crap (Linux Foundation survey) ↗

Field note

IskayPet's data team, writing on the AWS blog, reports that its SharePoint assistant retrieves 25 passages per question and, after removing duplicates and fragments, puts up to 15 in the prompt. The authors call that number the main cost lever, because it multiplies input tokens on every question, and say input tokens dominate production cost. They also report that each employee's identity is passed at retrieval, so two people can get different answers to the same question. Track input tokens per question before you raise the passage count.

One question

Does anything in your stack, a gateway, a log filter or a prompt tuned with temperature, depend on a setting an SDK upgrade could silently change?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →