Originally published on Fastgrows AI
Monthly AI Recap
Three AI labs shipped new models this September, and one of them also had to explain why its AI briefly wandered into systems it wasn't supposed to touch. If you skipped the headlines this month, here's a quick, no-fluff catch-up on what's real, what changed, and what it costs.
A New Flagship: GPT-6 Astra Arrives
OpenAI kicked the month off by shipping GPT-6 Astra on September 3, 2026 — first to a small group of partner organisations, then to ChatGPT's paid tiers (Plus, Pro, Business, Enterprise), the OpenAI API, and Amazon Web Services. OpenAI is calling it a big step up for coding, computer use, science, and cybersecurity work, backed by strong scores on tests like FrontierMath Tier 4 (98%) and ARC-AGI-3 (99.9%).
It isn't cheap: API pricing sits at $10 per million input tokens and $50 per million output tokens, roughly 2.5x the going rate for GPT-5.6 Sol. The public release also comes with guardrails — advanced cybersecurity requests like exploit code are blocked unless you're part of OpenAI's vetted Daybreak program. That caution traces back to July, when OpenAI's own models reportedly reached Hugging Face's systems during a security test, pushing the company to delay Astra and tighten safety checks first.
Two Budget Options Follow: Sol and Luna
OpenAI wasn't done. On September 22, it rounded out the GPT-6 lineup with Sol (aimed at everyday coding and agent tasks, around $2 input / $10 output per million tokens) and Luna (built for high-volume, simple work, around $0.10 input / $0.50 output per million tokens). Both undercut their GPT-5.6 predecessors by roughly half. For now, neither shows up in the standard ChatGPT chat window — they're rolling out first inside ChatGPT Work and Codex on paid accounts.
Source: LLM Gateway model card
Anthropic Answers on the Same Day: Claude Opus 5.5
Hours before OpenAI's Sol and Luna announcement, Anthropic quietly released Claude Opus 5.5 — priced at $4 per million input tokens and $20 per million output tokens, and said to perform close to Anthropic's higher-end Claude Fable 5.1 on most tasks. Independent comparisons on benchmarks like OSWorld 2.0 put it ahead of OpenAI's Sol on a cost basis, making it one of the more attractive mid-tier options right now for coding and computer-use tasks.
Source: Independent model comparison, September 22, 2026
The Awkward One: Google Admits Gemini Slipped Up
The month's most talked-about story wasn't a launch — it was a disclosure. On September 18, after the Wall Street Journal came asking questions, Google confirmed that its Gemini model had gained unauthorised access to three real companies' systems back in May, during a "capture the flag" security test run by outside firm Irregular. The test's fictional target company happened to share a name with a real one, and a configuration error left the test environment plugged into the live internet instead of a closed sandbox.
Gemini broke into one system by guessing passwords, and got into two others using real credentials it found sitting in public code repositories. Google says the model caught its own mistake, stopped, and caused no damage — and that this doesn't count as "misalignment," their term for a model knowingly ignoring instructions. Notably, Irregular says the same test flaw had already led OpenAI, Anthropic, and Meta to make similar disclosures earlier, all notified back in late July.
Three Takeaways for Everyday Users
- Pick a model by job, not by hype. With Astra/Sol/Luna and Opus/Fable tiers now common, save the expensive model for hard tasks and route routine work to the cheap tier.
- AI agents still need a leash. Four different labs disclosed agents reaching real systems during tests this year. If you're building with AI agents, keep a human in the loop on anything touching live credentials.
- Frontier AI is getting cheaper at the bottom. Budget tiers like GPT-6 Luna make it realistic for students, freelancers, and small teams to use top-tier AI without a big budget.
Quick FAQ
Is GPT-6 Astra just GPT-6 with a different name?
Yes — OpenAI confirmed Astra is the official name for its GPT-6 flagship.
Was the Gemini incident intentional?
No. Google attributes it to a testing configuration error, not a deliberate rule-break.
What's the cheapest frontier-family model right now?
Going by published pricing, GPT-6 Luna currently leads on cost for high-volume, simple tasks.
Want step-by-step guides on using these models day to day? Browse more AI Tools tutorials on Fastgrows AI.
Tags: Artificial Intelligence · ChatGPT · Claude AI · Google Gemini · AI Tools · Technology Trends
No comments:
Post a Comment