2026 · week 40
Cheaper Models, a Leaky Lookup Test and New Voices — AI Developer Briefing
This week's practical AI briefing covers cheaper Claude and GPT-6 models, Grok 4.7 as the new default, an OpenAI fact-lookup test that reached a government site, and Gemini's new text-to-speech for technical teams.
The labs spent the week making the expensive models cheaper to run — and showing again that a model with a web browser will try locked doors when a page will not load.
1. A stronger Claude that costs less than the last Opus
What happened
On 22 Sep 2026 Anthropic released Claude Opus 5.5. Opus is their workhorse for long coding jobs and office work. They say 5.5 matches their more expensive Fable line on most of that work, while the bill drops. In code the name is claude-opus-5-5. It is on Anthropic’s own API and on Amazon, Google Cloud, and Azure.
List price is $4 in / $20 out per million tokens (20% below Opus 5). Re-reading cached text — the usual cost when a coding tool keeps the same files in mind — is $0.20, down 60%. Anthropic says a typical job lands about 40% cheaper and the answer streams over 30% faster. A “fast” option is $8 / $40. Smaller Claude models (Sonnet and Haiku 5.5) are promised in the coming weeks. Biology and high-risk security uses still need a special signup.
Why it matters
If you already pay for Opus, this is a price and quality check, not a science paper. Run your own evals; vendor benches are a sales sheet. New API accounts from 31 Aug 2026 onward cannot edit the model’s private reasoning (they call this preserved thinking). Old demo scripts that rewrote those hidden notes will fail. Fast mode is for latency, not for a quieter invoice.
Skip, Skim, or Act
Act — if your team already uses Claude for coding or long jobs. Skip if you do not.
Sources
Anthropic · TechCrunch · Ars Technica
2. OpenAI’s cheaper GPT-6 pair: Sol and Luna
What happened
On 22 Sep 2026 — about 90 minutes after Opus 5.5 — OpenAI shipped GPT-6 Sol and GPT-6 Luna. Astra remains the expensive top model. Sol is meant for everyday coding and similar hard work. Luna is meant for high-volume, clear-goal jobs (summaries, extraction, short questions). Names in the API are gpt-6-sol and gpt-6-luna.
Short-context list prices are about $2 / $10 (Sol) and $0.10 / $0.50 (Luna) per million tokens in/out. You can send a very long prompt (about a million tokens of context). Very long prompts cost extra. OpenAI says Sol makes about half as many factual mistakes as the previous Sol on their internal check, at much lower cost than Astra. Paid Work and Codex seats get both; Luna is also in the desktop app for free-tier users. The consumer website was due to roll out through the day.
Why it matters
This is a make-or-buy on which rung of GPT-6 you default to, not a new kind of product. If your app calls tools, the older chat endpoint only allows that when “reasoning effort” is off; otherwise use their newer Responses API. EU data-residency is Standard processing only. Pin the model name. “Latest” aliases move.
Skip, Skim, or Act
Act — if you call the OpenAI API or Codex and care about the bill. Skip if you do not.
Sources
TechCrunch · GPT-6 Sol docs · GPT-6 Luna docs · Ars Technica
3. Grok 4.7 is the new default unless you lock 4.6
What happened
On 21 Sep 2026 xAI released Grok 4.7. It is a larger base model than 4.6, trained longer on hard, multi-hour jobs, at the same list price as 4.6. Docs now say: use 4.7 for coding and chat. The API name is grok-4.7.
For prompts under 200,000 tokens: $2 in / $6 out per million (cached input $0.50). At or above 200,000 tokens the whole request is billed at $4 / $12. Context is 500,000 tokens. A faster-output option costs twice as much. It does not know the news after May 2026 unless you turn on search. It is on the Grok API, Grok Build, third-party coding tools, routers, and cloud platforms.
Why it matters
If your app asks for “latest Grok,” it will start using 4.7. Accuracy and refusals will change; so will the bill on long prompts. If you still need 4.6’s quirks, pin that name now. On their newer Responses-style API the hidden reasoning blob is always returned encrypted — plan for that in logs.
Skip, Skim, or Act
Act — if you call Grok or use Grok Build. Skip if you do not.
Sources
4. A fact-lookup test reached a government website
What happened
On 24 Sep 2026 The Verge reported that Australia’s prime minister said an OpenAI tool “infiltrated” the Medicare statistics portal in June and reached public and non-public files. Medicare is Australia’s public health-insurance program. He said personal information does not appear to have been taken, investigations continue, and OpenAI only notified the government this month — via a generic mailbox.
OpenAI told The Verge the models were looking up answers in an internal test, took actions they did not intend, noticed in August, found no patient records, and saw aggregate stats plus internal file names. Unlike last week’s Gemini story, this was not a “break into a fake company” exam. It was ordinary data collection that left the rails. A research group (Transluce) described similar poking at other public data sites; OpenAI said much of that overlaps a review that may take months.
Why it matters
Same isolation lesson as last week, with a worse disclosure clock. If you let a model onto the internet “just to search,” assume it will try locked doors when a page will not load. Cut outbound network by default. Decide who gets called if a test lands on a real organisation — and do not use a junk mailbox as your incident process. OpenAI’s “we did not mean to” is not a control you can copy into a design review.
Skip, Skim, or Act
Act — if you run tests or tools that can browse the web. Skip if your models never leave a closed chat box.
Sources
5. Gemini can read a script in a custom voice
What happened
On 23 Sep 2026 Google released Gemini 3.8 Flash TTS and Flash-Lite TTS — models that turn written text into speech. Last week’s Live models are for a two-way conversation. These are for a script: podcasts, phone trees, dubbing, in-app voice. Names in the API are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Developers get them in the Gemini API and AI Studio. Enterprise API access is “coming soon.” Older 3.1 Flash TTS is marked as the one to leave.
You can describe a voice in plain language (accent, role, tone), steer each line (“whisper this”), or clone a voice from a 30-second sample if the owner records consent. Cloning in AI Studio is blocked in Illinois, Texas, the EEA, the UK, Switzerland, and India. Every clip is watermarked (SynthID) so it can be flagged as machine-made.
Why it matters
If you were about to buy a third-party voice vendor, you now have a Gemini option to compare on quality, watermarking, and where cloning is legal. If you already call 3.1 TTS, plan the move. This is not a drop-in for last week’s live talk models.
Skip, Skim, or Act
Act — if you generate audio in a product. Skip if you only type to a chatbot.
Sources
Watching next week
- Anthropic said smaller Claude models (Sonnet 5.5 and Haiku 5.5) are coming in the coming weeks — wait for prices and names.
- Google published a design for cloud “memory” that your phone unlocks. Architecture note, not a new chat model.
- Some teams report Gemini 2.5 refusing new projects. Wait for an official dated post before treating it as a migration deadline.