Tech • AI • Robotics • Game
Huge OpenAI DevDay Leak! “o” AI Agent, Sonnet 5.5 Beats GPT-6, MiniMax M3.1 Out & More! AI News
9.4/10TL;DR
OpenAI appears poised to use its September 29 Dev Day to unveil an always-on agent called O and broaden access to an ultra-fast API, as rivals including Anthropic and MiniMax push new models and coding capabilities.
KEY POINTS
OpenAI’s leaked agent project
References inside ChatGPT code and upgrade screens point to a product named O, described as an always-on assistant. Internal clues suggest support for 63 languages, multiple paid tiers, and possible links to email-style task handling. The system is reportedly aimed at long-running jobs, potentially operating for hours, days, or weeks rather than returning a response in minutes.
AON and long-horizon automation
The assistant is also tied to the codename AON, which appears associated with agent infrastructure for extended workflows. Reports indicate it could run in a hosted cloud environment, execute code, do research, and continue working until an objective is completed. Rumors also suggest multi-agent coordination, a design that would place it alongside emerging agent products such as Rabbit, Manus, and Hermes Agent.
Ultra-fast API expansion
A new selector spotted in the Responses API playground shows standard, fast, and ultra fast modes. Ultra fast had previously been limited to select customers, with claims that GPT-5.6 Soul could generate up to 750 tokens per second and run as much as 14 times faster than standard service. Its appearance in documentation as an access-controlled tier suggests a broader rollout may be near.
Speed becomes a product strategy
The API changes align with signs that OpenAI is increasingly selling latency as a premium feature. A recently spotted $500 ChatGPT Pro Max plan reportedly promises the fastest responses, indicating a tiered model where customers pay for inference speed as well as model quality. That would make infrastructure and responsiveness central themes of the company’s developer strategy.
Hints of additional OpenAI updates
Senior product messaging has raised expectations for a major event, with suggestions that more system “resets” or refreshes are coming next week. While details remain unclear, speculation has centered on a possible new model linked to the broader Astra family. Even without confirmation, the timing has reinforced expectations that Dev Day will include both product launches and backend changes.
Anthropic’s Sonnet 5.5 nears release
Anthropic has confirmed that Sonnet 5.5 is due within weeks, and reports say partners have already tested a newer checkpoint. Early descriptions portray it as very fast, efficient, and more natural in conversational style. If the model arrives near current Sonnet 5 pricing, it could become a major pressure point for OpenAI in coding and general-purpose use.
Price-performance pressure
Rumored pricing for Sonnet 5.5 is $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache reads, though those figures are not confirmed as final. The benchmark to watch is Opus 5.5, which Anthropic said became more than 30% faster while costing 40% less on typical workloads than Opus 5. A similar improvement at Sonnet pricing would make the model highly disruptive.
MiniMax launches M3.1 Flash Preview
What first appeared in an early GitHub commit has now become an official release: MiniMax M3.1 Flash Preview. The company is positioning it as a text model for everyday software development, emphasizing speed, reliability, and support for tasks from bug fixes to full feature building. The quick shift from leak to launch underscored how rapidly the model market is moving.
Fast models are closing the quality gap
Early image-generation comparisons suggest M3.1 Flash performs strongly for a speed-focused model. In one SVG test involving Super Mario and Peach, it did not clearly surpass a larger premium model, but it produced a coherent, detailed result that was competitive enough to draw attention. That kind of output suggests lighter, cheaper models are improving faster than many developers expected.
Long-running coding agents are becoming practical
A standout example from the broader market showed an advanced agent building a Minecraft-like project over about 25 hours. The output reportedly included roughly 100,000 lines of game code, 26,000 lines of tests, 1,000 automated tests, 775 files, 783 images, and 184 sounds, plus an in-game guide. The scale illustrates how AI coding agents are moving beyond short demos into sustained software production.
CONCLUSION
The immediate contest in AI is no longer just model intelligence but also speed, price, and autonomous execution. If the latest leaks and launches hold, Dev Day could mark a shift toward persistent agents and infrastructure as the next battleground.
Ask a question