FORSMILE
Issue #6Published September 27, 2026 / Covering Sep 22-27日本語で読む

This Week in AI, 8 Items — OpenAI and Anthropic Both Cut Prices on 22 September, Moving the Contest to Cost per Task

The week in one place

The axis this week was price. **On 22 September, OpenAI and Anthropic shipped flagship models on the same day, and both cut what they charge.** OpenAI halved GPT-6 Sol and Luna against GPT-5.6 pricing (Sol $4 to $2 in, $20 to $10 out; Luna $0.20 to $0.10 in, $1.20 to $0.50 out). Anthropic priced Opus 5.5 at $4 input and $20 output, 20% under Opus 5, with cache reads at $0.20, 60% under Opus 5. What is notable is that the figures each company led with were **cost per task**, not headline capability: OpenAI published AutomationBench results of 33.2% at $0.27 per task for GPT-6 Sol against 26.9% for Claude Opus 5 (max) at 11.1x Sol’s cost, while Anthropic said Opus 5.5 runs 40% cheaper than Opus 5 on typical workloads. **And the Opus 5 that OpenAI benchmarked against was superseded by Opus 5.5 the very day OpenAI published** — a concrete demonstration that a vendor comparison table can be a generation stale on the day it goes up. The other thing the two shared was vocabulary. Anthropic called Opus 5.5 "our first release since we called for pacing the frontier"; OpenAI framed its third-party assessment principles as "part of our efforts to pace the frontier". Cost and external scrutiny moved forward in the same week.

Models & Products

Anthropic ships Claude Opus 5.5 at $4 / $20 with $0.20 cache reads, and says it runs 40% cheaper than Opus 5 on typical workloads

The first model in the new Claude 5.5 family, cheaper and faster than Opus 5.

Published 22 September. Anthropic says it "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5". Input and output are $4 and $20 per million tokens, 20% less than Opus 5, and **cache reads are $0.20 per million, 60% less than Opus 5**. Output generation is more than 30% faster than Opus 5. The Claude Code changelog records `claude-opus-5-5` becoming the default Opus model in 2.1.280, with a **1M context window**. Because it is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with safeguards similar to those on Claude Fable 5.1. Claude Sonnet 5.5 and Claude Haiku 5.5 are said to follow "in the coming weeks".

So What

The 60% cut on cache reads is the part that shows up in a bill. Anthropic itself notes that cache reads "make up the majority of agentic and coding work costs", so the same workflow gets cheaper without being rewritten. Note that **both "the level of Fable 5.1" and "40% cheaper" are Anthropic’s own measurements**; what is externally attested is only that Frontier Design and METR tested the model before release.

Models & Products

OpenAI ships GPT-6 Sol and Luna, halving API prices against GPT-5.6

Two cheaper models below Astra, with input and output prices cut 50%.

Published 22 September. Both were trained with methods similar to GPT-6 Astra, aimed at the cost-efficiency end of the family. Per million tokens, **Sol goes from $4 to $2 in and $20 to $10 out**, and **Luna from $0.20 to $0.10 in and $1.20 to $0.50 out**; OpenAI describes this as 50% off GPT-5.6 promotional pricing. On AutomationBench 1.0.6, which tests agents on end-to-end business workflows using 47 tools, GPT-6 Sol (xhigh) scores 33.2% at $0.27 per task, Claude Opus 5 (max) scores 26.9% at 11.1x Sol’s cost, and Claude Fable 5.1 with Opus 5 fallback (max) scores 31.4%. On Agents’ Last Exam V1, Sol (max) scores 56.4%. The models are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu; Free and Go users get Luna in the desktop app. **They are not in Chat yet.** API ids are `gpt-6-sol` and `gpt-6-luna`.

So What

**What dropped is the API token price.** Anything you run through the API gets cheaper per unit. OpenAI does not say that Codex subscription pricing or credit-denominated cost fell, so **this announcement is no basis for expecting a flat-rate bill to drop.** Read the comparison table carefully too: **the Claude Opus 5 that OpenAI benchmarked was superseded by Opus 5.5 on the day this post went up.** OpenAI also notes that competitor scores were taken from public reports, and that Fable 5 numbers were used where Fable 5.1 numbers were unavailable. Treat a vendor table as stale from publication and measure on your own workload.

Developer Tools

OpenAI improves GPT-6 prompt caching: 90% off cached input reads, and changing reasoning effort or tool availability no longer breaks the cache

Higher default cache hit rates, plus controls that preserve earlier context for reuse.

Published 22 September. OpenAI raised default cache hit rates for GPT-6 and gives **a 90% discount on cached input-token reads**. Three things came with it: (1) a Prompt Caching Dashboard and a diagnostics tool that show how much input is cached and explain missed caching opportunities; (2) **raising or lowering reasoning effort, and enabling or disabling tools, now preserve earlier context for cache reuse**; and (3) explicit breakpoints let developers choose where a cached prompt prefix ends. OpenAI cites GitHub reporting that these improvements cut the share of prompt tokens needing fresh processing by more than 50% across billions of requests.

So What

This matters when running agents. Previously, raising reasoning effort mid-task invalidated the cache and both cost and latency jumped; that constraint is gone. **It changes a design decision** — varying effort by task difficulty becomes practical rather than expensive. The 50% figure is GitHub’s own reporting, quoted by OpenAI.

Developer Tools

Claude Code 2.1.280 to 2.1.283: Opus 5.5 becomes the default Opus, plus settings that pin models by version and an audit for prompts written for older models

Controls to stop new models being used automatically, and a command that finds stale CLAUDE.md patterns.

**In 2.1.280 (22 September)**, `claude-opus-5-5` became the default Opus model (1M context, $4/$20 per Mtok, $0.20/Mtok cache reads). **In 2.1.283 (25 September)**, setting **`availableModelsMatch` to `"exact"` means an `availableModels` entry allows only the version it names, so new releases stay blocked until listed**, and **`deniedModels`** blocks specific models outright. The same version added **`/doctor prompt-audit`** (also `/checkup prompt-audit`), which audits CLAUDE.md files, skills, agents and commands for prompting patterns written for older models. **2.1.282 (24 September)** added `maxProseWidth` to cap prose width in wide terminals, and **2.1.281 (23 September)** added `"attribution": false` in `settings.json` to hide all commit and PR attribution.

So What

**Two vendors shipped models this week, which makes model pinning immediately relevant** — you can hold a version until your own evaluation finishes. Separately, `/doctor prompt-audit` earns its keep in proportion to how many skills and CLAUDE.md files you carry. **The official docs changelog does date each version, and all four fall inside this window** (2.1.280 on 22 September, 2.1.281 on the 23rd, 2.1.282 on the 24th, 2.1.283 on the 25th). 2.1.280, which made Opus 5.5 the default Opus, landed on 22 September — the same day Anthropic shipped the model.

Regulation & Safety

OpenAI publishes four priority areas and principles for third party safety assessments, and defines "safety claim" and "safety case"

A written policy on what external assessors get access to, and how deep.

Published 22 September. Framed as "part of our efforts to pace the frontier", OpenAI commits to supporting independent assessments with deep access across training, evaluation and deployment. The document defines a **safety claim** as a specific assertion about a model’s capabilities, behaviour or safeguards that can be assessed against evidence, and a **safety case** as a structured, evidence-backed argument connecting individual claims and making assumptions, uncertainties and residual risks explicit. It then sets out **four priority areas** for deeper assessment alongside principles. This work is long-term and launch-agnostic, expected to run **from weeks to several months**, with multiple assessments in parallel. Access already provided includes information on technical safeguards, visible chain-of-thought access, and confidential data and internal deployment access for incident response and monitor red teaming.

So What

Useful as vocabulary if you are ever on either side of an external audit. Separating "claim" from "case" transfers directly to vulnerability write-ups and security documentation. **This is a statement of policy, not a report of what any assessor actually found.** Read alongside Anthropic commissioning METR for an independent investigation, covered in issue 5: both labs are now publishing about external scrutiny.

Market & Industry

ChatGPT Ads reaches six Southeast Asian markets and Taiwan, now over 60 countries, still Free and Go only

Ads roll out in Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan.

Published 23 September. ChatGPT Ads begins rolling out in Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, taking it to **more than 60 countries**. It follows Asia Pacific launches in Australia, New Zealand, Japan, South Korea and India. Advertisers buy through the OpenAI Ads Solutions team, agency partners including dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP, and technology partners; eligible businesses can self-serve through Ads Manager. **Ads are shown only to users on the Free and Go plans; Plus, Pro and Enterprise stay ad-free.**

So What

Tracked less as a place to buy ads than as **a change in the shape of traffic**. The more countries where the free tier of ChatGPT carries advertising, the more ad inventory exists upstream of the answer rather than in search results. Japan is already covered, so this is context for reading a gap between flat impression counts and real demand. **The announcement gives no data on impressions or revenue effects**, so there is nothing causal to claim yet.

Research

Anthropic sets up a life sciences research group and lab; Claude found a novel enzyme system with CRISPR-like repeat arrays

About 950 agents spent 21 hours and 210 million tokens and surfaced one uncharacterised system.

Published 23 September. Anthropic formed an in-house life sciences research group and wet lab, and reports as an early result that Claude found **a novel enzyme system associated with an array of DNA repeats**. It is built on a **reverse transcriptase (RT) found in a jumbo phage**; the RT itself had been identified in earlier studies, but **Claude appears to be the first to notice the defining features — an associated array of non-coding DNA sequences and an additional accessory protein of unknown function**. Anthropic calls these array-associated reverse transcriptases (ART). Human involvement was limited to the initial prompt and the lab work; **roughly 950 agents searched the sequence database for 21 hours using 210 million tokens**. Feng Zhang of MIT and the Broad Institute, a CRISPR pioneer, reviewed the pre-print and called the finding "genuinely intriguing" and worth further investigation. **The primary function of ARTs is not yet known and the work is ongoing.**

So What

Valuable as a worked example of massively parallel agent search with actual numbers attached. **950 agents, 21 hours, 210 million tokens** gives a yardstick for how many orders of magnitude any batch you run sits below this. Read the result itself carefully: the function is unresolved, the RT was already known, and the outside comment goes only as far as "merits further investigation". **Summarising this as "AI discovered a new enzyme" drops every one of those qualifiers.**

Models & Products

Google adds 14 Connected Apps to Gemini, callable with an @ mention

Airtable, Linear, Adobe, Webflow and others can now be driven from inside Gemini.

Published 23 September, rolling out from that date. Productivity: Airtable, Linear, monday.com, PandaDoc, Wispr AI and Zoho. Creativity: Adobe, Picsart, Squarespace and Webflow. Lifestyle: apartments.com, Experian for credit monitoring, Peloton and SeatGeek. You connect them in Gemini settings, or pull them into a chat with an **@ mention** or by asking directly.

So What

Only Linear and Webflow touch day-to-day work here: issue tracking and site building are now reachable from the chat surface on the Gemini side too. **This is an announcement that more destinations exist, with nothing said about permission granularity or audit logging.** Whether to connect real business data is a decision to make after that information appears.

Watching (no confirmed primary source yet)

  • The window starts on 22 September so it joins cleanly to issue 5, which covered 14 to 21 September, with no overlap and no gap. The 26th and 27th were a weekend with no vendor announcements; OpenAI’s news index had nothing later than 23 September.
  • **Closing the "California SB 1047 signature deadline of 30 September" item from issue 5: the premise was wrong.** SB 1047 was vetoed by Governor Newsom in September 2024 and never became law, so there is no 2026 signature deadline. Its successor, SB 53 on frontier AI transparency, passed in 2025 and took effect on 1 January 2026. Issue 5 correctly declined to run the item for lack of primary sourcing, but **the reported premise itself was false**, so this thread is dropped rather than carried forward.
  • **The Meta, Amazon and Microsoft retrieval problem noted in issue 5 is resolved by changing route.** The ai.meta.com blog index still cannot be fetched directly, but domain-scoped search into individual pages works. Doing that established that Meta’s personal agent Muse was announced on **8 September**, outside this window and already covered in issue 4. No in-window announcement from Meta, Amazon or Microsoft was confirmed.
  • Aggregator articles report that Meta shipped a security hotfix for its new agent, that Shopify opened a checkout lane Amazon had closed, and that China is investigating its own model developers. **None could be confirmed from a first-party announcement, so none is included.** They will be covered once they can be.
  • Google published a piece on Project Suncatcher, its effort to put AI compute in space, on 24 September, but **the substance is the release of a video series about an existing project**, and most of the page body is a Google AI generated summary. No new fact could be extracted and no `soWhat` written, so it is not an item.
  • OpenAI DevDay 2026 falls on 29 September in San Francisco, outside this window. It goes in the next issue.
  • **This issue originally stated that the Claude Code changelog carries no dates. That was wrong.** The CHANGELOG.md on GitHub has none, but once the citation was switched to the official docs changelog, that page does date each release: 2.1.280 on 22 September, 2.1.281 on the 23rd, 2.1.282 on the 24th and 2.1.283 on the 25th — **all four inside this window** (2.1.278 is dated 19 September, in issue 5’s window). **When the source changes, rebuild the claim from the new source.**

"Primary" links go to the announcing party's own publication (company blog, press release, official docs). "Reporting" links go to news coverage or third-party analysis. Figures and dates are as verified on the publication date.

← All Weekly AI News issues