GPT-6 Astra Pro now supports multiple agents working on tasks in parallel. The Claude Code team reports that Claude writes about 90% of its code; merged PRs more than doubled while the rollback rate fell by over 40%. Reflection has released Beam 1.0 and its training tools, offering a 1-million-token context window under the Apache 2.0 license; Step-5-Preview, which also targets long-context tasks, uses a non-commercial license. The Ministry of Commerce and the Ministry of Science and Technology propose adding 23 AI technologies to export controls. Exports of the listed technologies must be registered before proceeding, effective from the announcement's publication, but API and SaaS services fall outside the scope of large-model technology export controls.
Dividing tasks boosts output, while continuous review keeps the risks of growing code changes in check.
OpenAI has upgraded GPT-6 Astra Pro in ChatGPT to let multiple agents work on different parts of the same task in parallel. Users can track each agent's progress and open completed tasks to view the results. The upgrade also speeds up responses and improves the use of plugins and tools.
Read original →The Claude Code team reports that Claude writes about 90% of its code. From June to September, merged PRs more than doubled, while the rollback rate fell by over 40%. Review has shifted from a manual checkpoint after submission to a continuous process: Claude reviews every PR first, and most changes have been checked and fixed at least once before a human gets involved. The team uses rollback and incident rates as guardrails to avoid turning time saved on reviews into time spent fixing failures.
Read original →🔍 Analysis: GPT-6 Astra Pro divides a single task among multiple agents, while Claude Code's experience addresses another question: can checks keep pace as generation capacity grows? The Claude Code team more than doubled merged PRs while reducing its rollback rate by over 40%, showing that higher output did not come at the cost of more rollbacks. When evaluating parallel agents, enterprises should measure validation and remediation capabilities alongside task throughput, rather than simply counting how many agents can run at once.
The long windows in Beam, Step 5, and Gemini 4 Argon need to be assessed separately for usage rights and capacity limits.
Reflection has released Beam 1.0, an MoE model with 501 billion total parameters, 23 billion active parameters, and a 1-million-token context window. The model is available under the Apache 2.0 license, alongside weights in three precision formats, the Mirror CLI tool used during training, and a safety report. Safety training addresses both harmful requests and excessive refusals, aiming to reduce harmful assistance while preserving the ability to answer legitimate questions on sensitive topics.
Read original →StepFun has released open weights for Step-5-Preview, which uses an MoE architecture with 600B total parameters and 27B active parameters. The model accepts text and image inputs and produces text output, with context and output limits of up to 1M tokens, targeting repository-level software engineering and long-horizon tool use. Unlike the earlier StepFun Flash model, which used Apache 2.0, this preview uses a non-commercial license.
Read original →Google AI Pro subscribers can now use Gemini 4 Argon in AI Studio's Build and Playground. This expanded access raises the output limit from 64K to 1 million tokens, removing the previous 64K cap. However, input and output share the same window: the million tokens are not an additional allowance separate from the input.
Read original →🔍 Analysis: Beam's 1-million-token context window, Step-5-Preview's 1M-token context and output limits, and Gemini 4 Argon's increased output limit cannot simply be treated as equivalent capacities for long tasks. Argon shares a window between input and output, so long inputs leave less room for generation. Deployment choices face another constraint: Beam uses Apache 2.0, while Step-5-Preview uses a non-commercial license. For enterprises, a sufficiently long window does not mean the models can be put to commercial use in the same way.
From adjustable reasoning to recurrent visual computation, shrinking models is no longer the only route to lower costs.
Anthropic has released Claude Haiku 5.5, saying its average running cost is about 75% lower than Haiku 4.5's, targeting high-frequency tasks such as summarization and database queries. It introduces adjustable reasoning effort to the Haiku series for the first time, allowing users to balance cost and capability, and can also serve as a coding subagent alongside Opus 5.5 and Sonnet 5.5. It is now available on Claude Platform, AWS, Google Cloud, and Azure.
Read original →Elorian has released the vision-language model Glimpse, which is now available to try. Rather than relying on a text-only chain of thought, it performs recurrent computation over latent visual states. Elorian says the model matches or exceeds Astra on 5 of 12 benchmarks, with token pricing at roughly 1/17 of Astra's, and was trained using hundreds of GPUs.
Read original →Perplexity has released the open-weight decision model pplx-decider v1.2 in 27B and 4B Flash versions. The update increases training compute and dataset size, and adds reranking and multi-question reasoning in a single pass, with the latter offering speedups of up to 5×. The two models cost 2 cents and 1 cent per million tokens, respectively.
Read original →🔍 Analysis: Haiku 5.5 matches adjustable reasoning effort to task difficulty, Glimpse performs recurrent computation over latent visual states, and pplx-decider v1.2 offers specialized capabilities for decision-making, reranking, and multi-question reasoning. Together, they point toward allocating inference resources by task rather than using the same model tier for every request. Their cost figures cannot be ranked directly, either: Haiku's roughly 75% figure is an average running-cost reduction versus its predecessor, Glimpse's roughly 1/17 figure compares its token price with Astra's, and pplx-decider quotes prices per million tokens. For developers, model selection should come down to performance and completion cost on specific tasks, not just the size of the price cut.
Local deployments overseas and API or SaaS services fall on different sides of the regulatory boundary.
On October 16, 2026, the Ministry of Commerce and the Ministry of Science and Technology sought public comment on revisions to the technology export catalogue, proposing to add 23 AI technologies, including large models, agents, and computing clusters. Exports of the listed technologies must be registered before proceeding, effective from the announcement's publication. The proposed rules cover local overseas deployments of restricted technologies, but services such as APIs and SaaS fall outside the scope of large-model technology export controls. Technologies that have entered the public domain through open source or published papers would be excluded from controls, and exports of open-source technologies could qualify for facilitation measures such as automatic licensing.
Read original →🔍 Analysis: The proposed rules bring local overseas deployments of restricted technologies within scope, while excluding APIs and SaaS from large-model technology export controls. This means the way a capability is delivered affects the regulatory process that applies. For companies expanding overseas, technology delivery and service delivery cannot be assessed as one category. Meanwhile, the catalogue revision covering 23 technologies is still open for public comment, while the prior-registration requirement took effect upon publication of the announcement, creating a separate timeline from the rules' eventual finalization.