AI NewsSeptember 28, 2026

Big announcements often land at this event, but none are confirmed yet. Here is what OpenAI has actually said about the schedule and access.
OpenAI's annual developer conference takes place September 29 at Fort Mason in San Francisco. CEO Sam Altman opens with a keynote at 10 a.m. Pacific. In-person spots were application-only and are now closed.
Any model or product news is still speculation. Past DevDay keynotes have carried major API announcements, so developers will be watching closely. If you miss it live, recordings are posted afterwards.
Key Capabilities:

• Free keynote: Altman's 10 a.m. Pacific talk streams live for anyone, no ticket needed.
• Ticket cost: In-person registration for invitees was $650, with scholarships available.
• Global tour: DevDay Exchanges run in Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City.

Presented at IROS 2026, SAIL is a joint project with the University of Tokyo. A robot's AI model tests and refines movements in simulation before acting.
Robots often fail when given one shot at a new task. SAIL lets them rehearse first, like a chess player thinking ahead. Across six simulated tasks, success rose from 25% to 73%.
The AI model is not retrained. It just gets more thinking time before moving, which needs a simulator. The payoff: robots that fail less without piles of new human demos.
Key Capabilities:

• Real robot test: On a physical SO-101 arm, one task succeeded in five of six trials.
• Up to 95%: Success on complex tasks reached 95% as the search budget grew.
• Under the hood: It uses Google's Gemini Robotics-ER 1.5 as both the planner and the judge.

A new slash command makes this AI coder research first and wait for a green light before editing files. It mirrors the planning mode already in Antigravity's command-line tool.
On September 26, Google announced /plan for Antigravity 2.0. Type it and the agent researches your codebase, then writes a step-by-step plan for review. It normally waits for your approval before touching any code.
Developers reacted with mixed feelings. A May forum thread asked where the IDE's plan mode had gone, and some mocked the return as a step backward. Still, a plan you can approve first means fewer surprise edits to your files.
Key Capabilities:

• Light version: Asking for a plan in a normal prompt gives a shorter, lighter plan.
• Your call: Unless your review policy is set to Always Proceed, you approve before execution starts.
• Edit the plan: You can comment on steps or ask the agent to shrink the scope

Open weights, an MIT license, and a claim that AI helped build it. Released September 27, Naive-N0.5-Flash targets coding and AI research.
Reading a million tokens (about 750,000 words) at once normally demands huge memory. NaiveAI's model skips the costly step of comparing every word to every other, like skimming with a highlighter. Anyone can download it.
Benchmark claims are the company's own. NaiveAI, based in Beijing, says AI wrote code and ran experiments during development, with humans setting goals. If true, AI is now helping build the next AI.
Key Capabilities:

• Big but light: 309B parameters, but only 15.5B active at a time, which keeps running costs down.
• Very fast: Its runtime claims up to 2,000 tokens per second in Ultrafast mode.
• Cheap API: The hosted API lists $0.10 per million input tokens and $0.40 for output.

An image-encoding flaw had been dragging down how both models read pictures. OpenAI shipped the patch across its developer platform and wants image-based evaluations rerun.
Launched September 22, GPT-6 Sol and Luna were misreading pictures. On a visual grounding test (finding objects in an image), OpenAI's chart shows Luna jumping from 28.8% to 60% after the fix.
Any image test run before the patch is now outdated. Teams that judged either model on screenshots or computer-use tasks should rerun those evaluations and retry affected workflows, as OpenAI recommends.
Key Capabilities:

• Big jump: Luna's visual grounding score rose from 28.8% to 60% in OpenAI's own chart.
• Wide reach: Better results now show up in the API, Codex, and computer-use tasks.
• Cheap tiers: Sol and Luna cost half as much as their GPT-5.6 counterparts.

Developers using MiniMax Code now have a lighter, quicker model option. It comes with five reasoning levels and no public price, benchmarks, or API yet.

On September 27, MiniMax added M3.1-Flash-Preview to its coding tool. It targets daily work, from quick bug fixes to full features. Users choose a reasoning level from low to max.

Details are thin. There is no model card, no price, and no public API. A free stealth model, Space Bunny Alpha, is reportedly an early build, but MiniMax has not confirmed that.
Key Capabilities:

• Reasoning dial: Five effort levels, low to max, let developers trade speed for depth.
• Bonus credits: MiniMax Code users get double daily check-in credits from September 28 to October 7.
• Limited access: Available only inside MiniMax Code for now, so no API calls yet.