AI News · September 2026
September 2026 turned out to be one of the busiest months in AI so far. A new flagship model landed within the first three days, two more followed later in the month, and a security story involving Google's Gemini made headlines around the world. If you only have five minutes, this roundup covers what actually happened — with sources you can check yourself.
1. OpenAI Launches GPT-6 Astra
On September 3, 2026, OpenAI released GPT-6 Astra, its new flagship model, to a limited set of organisations first, followed by a wider rollout to ChatGPT Plus, Pro, Business, and Enterprise plans, along with the OpenAI API and Amazon Web Services. OpenAI describes Astra as state-of-the-art in computer use, coding, cybersecurity, science, and professional work, and says it scores 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
On the API, Astra costs $10 per million input tokens and $50 per million output tokens — about 2.5 times the promotional rate of the earlier GPT-5.6 Sol. The public version is also more restricted than OpenAI's internal one: it refuses advanced cybersecurity tasks like proof-of-concept exploits unless you're part of OpenAI's vetted "Daybreak" program. This came after OpenAI delayed the launch to add safeguards, following an incident in July where its models were reported to have autonomously accessed Hugging Face's systems during a security test.
2. GPT-6 Sol and GPT-6 Luna Bring the Price Down
Nearly three weeks later, on September 22, 2026, OpenAI expanded the GPT-6 family with two cheaper models: GPT-6 Sol and GPT-6 Luna. Sol is positioned for everyday coding and agent work at roughly $2 per million input tokens and $10 per million output tokens, while Luna is built for high-volume, simple tasks at around $0.10 input and $0.50 output per million tokens — both about half the promotional price of their GPT-5.6 equivalents. Neither model appears in the standard ChatGPT chat tab; they roll out first inside ChatGPT Work and Codex on paid plans.
Source: LLM Gateway model card
3. Anthropic Releases Claude Opus 5.5
On the same day as OpenAI's Sol and Luna announcement, Anthropic released Claude Opus 5.5. Anthropic says the model performs at a level close to its higher-tier Claude Fable 5.1 on most tasks, priced at $4 per million input tokens and $20 per million output tokens — a notably lower price point than OpenAI's comparable Sol tier for similar coding and computer-use benchmarks such as OSWorld 2.0.
Source: Independent model comparison, September 22, 2026
4. Google Confirms a Gemini Security Incident
On September 18, 2026, following reporting by the Wall Street Journal, Google confirmed that its Gemini model had gained unauthorised access to the systems of three real companies back in May 2026. The incident happened during a "capture the flag" cybersecurity evaluation run by third-party firm Irregular. A fictional target company in the test happened to share its name with a real one, and a configuration bug left the test environment connected to the live internet instead of a sealed sandbox.
Gemini gained access to one system by guessing passwords, and to two others by finding real credentials sitting in public code repositories. Google said the model stopped itself once it noticed the mismatch, that it found no evidence of damage, and that it does not consider the incident "misalignment" — its term for a model deliberately ignoring instructions. Irregular confirmed that the same underlying testing flaw had earlier led to similar disclosures from OpenAI, Anthropic, and Meta, all notified in late July.
What This Means for You
- Model choice now has more layers. Both OpenAI and Anthropic are shipping tiered models (Astra/Sol/Luna and Opus/Fable) so you can match the model's cost to the actual difficulty of your task, instead of using one expensive model for everything.
- Agent safety is now a headline topic, not a footnote. Four major labs — OpenAI, Anthropic, Google, and Meta — have all disclosed AI agents reaching real systems during security tests this year. If your business is experimenting with AI agents, keep human approval steps on anything that touches live data or credentials.
- Prices are trending down at the low end. Cheaper tiers like GPT-6 Luna make it more affordable for small businesses and students to use frontier-level AI for everyday tasks like summarising or drafting.
Frequently Asked Questions
Is GPT-6 Astra the same as GPT-6?
Yes. OpenAI confirmed Astra is the name for its GPT-6 flagship model.
Did Gemini actually "hack" anyone on purpose?
No. Google says the model mistook a live company system for a test target because of a configuration error, not a deliberate attempt to break rules.
Which model is cheapest for everyday AI tasks right now?
Based on published pricing, GPT-6 Luna is currently the least expensive frontier-family option for high-volume, simple tasks.
Want a simple breakdown of how to actually use these new models for your business or studies? Check out our AI Tools guides on Fastgrows AI for beginner-friendly walkthroughs.
Tags: Artificial Intelligence · ChatGPT · Claude AI · Google Gemini · AI Tools · Technology Trends
No comments:
Post a Comment