OSC 7501
Oct 14, 2026

OSC 7501 gains support from Claude Code, Grok Build and more terminal tools

OSC 7501, a terminal state protocol proposed by Ghostty author Mitchell Hashimoto, is starting to see adoption across coding agents and terminal tools. Anthropic’s Claude Code v2.1.295 and xAI’s Grok Build v1.0.52 already support the protocol, while OpenAI has merged Codex adaptation code on GitHub, though it has not yet appeared in the latest stable release. The protocol lets an agent explicitly report whether it is working, waiting for user input, finished, or in an error state, giving terminals a direct signal instead of forcing them to infer status from loading animations or on-screen text. BlockBeats said several open-source agents have also added support, including Pi v1.1.0, Crush v0.98.0, Oh My Pi v18.8.7, and Docker Agent v1.150.0. OpenTUI, the terminal UI framework used by OpenCode, added an OSC 7501 interface in version 0.6.3, although OpenCode itself has not yet merged its adaptation code. On the terminal side, Rex, cmux, and TUIOS can already receive these state signals.

3430
OSC 7501 gains support from Claude Code, Grok Build and more terminal tools
OpenAI
Oct 13, 2026

OpenAI raises GPT-6.1 Sol default context window to 500k and drops long-context surcharge for subscribers

OpenAI has expanded the default context window for GPT-6.1 Sol in Codex and ChatGPT Work from 272k to 500k, an increase of roughly 80%, according to the update described in the source material. Tibo said users running GPT-6.1 Sol under subscription credits will no longer pay an extra fee for long context, while API usage will still be billed under the API pricing schedule. The change targets a common complaint from developers who use coding agents for longer jobs: repeated context compression, which can interrupt work, slow execution, and force users to restate earlier requirements. The report also notes that a larger window does not automatically reduce credit consumption. OpenAI previously raised context from 272k to 372k for GPT-5.6 Sol in July, then reverted the change after subscription usage was deducted faster than expected. Tibo said the issue was not a long-context surcharge but repeated reads of cached content. The article compares the new default with Claude Code, where Opus 5.5 and Sonnet 5.5 reportedly already run with a 1M default context window, while GPT-6.1 Sol can also reach 1M if users raise the setting manually.

610
OpenAI raises GPT-6.1 Sol default context window to 500k and drops long-context surcharge for subscribers
OpenAI
Oct 13, 2026

OpenAI resets global usage limits again on Day 8 of its 28-day sprint

OpenAI product and platform lead Tibo Sottiaux said the company’s Day 8 update in its 28-day product sprint did not include a new feature release. Instead, OpenAI announced another global usage-limit reset, which is expected to be completed before the end of Oct. 12 on U.S. West Coast time. The move comes just two days after a previous global reset on Day 6. As with the earlier reset, eligible usage quotas for Codex and ChatGPT Work will refresh automatically. Users do not need to claim the reset manually, and the refreshed quota cannot be carried over for later use. The update was reported in a newsflash by BlockBeats.

830
OpenAI resets global usage limits again on Day 8 of its 28-day sprint
OpenAI Math Drop
Oct 13, 2026

Open-source push on GitHub lifts OpenAI Math Drop problem #109 parameter to 0.00129 in about 24 hours

A fast-moving open-source effort around OpenAI Math Drop problem #109, which focuses on lower bounds for conditional integer multiplication, produced a sharp jump in the key parameter κ over roughly 24 hours. According to the report, work in the GitHub repository CrocSwap/integer-mult-bounds pushed κ from 2^-182 to 0.00129, with mathematicians, developers and AI agents contributing in parallel. The article says the repository received 151 pull requests in a single day and generated more than 260 million lines of testing and optimization code. Doug Colkitt made the early gains, moving the benchmark from 2^-182 to 2^-108, then 2^-78 and later 2^-34. After that, Rohan Arun and others deployed dozens of Codex- and Claude-based agents through automated research tooling including GetSupers, allowing code generation and verification to run at high speed. The report also highlights differences between the models. Of the 151 PRs, 89 were tagged with Codex and 48 with Claude. Codex Astra high was described as stronger in rapid iteration and physical cost optimization, while Claude Opus 5.5 was credited with some of the leading structural advances. A later step involved Jacob Sussman introducing an E8 lattice geometry approach, which the article presents as a turning point before the best reported result of κ = 0.00129 was submitted.

950
Open-source push on GitHub lifts OpenAI Math Drop problem #109 parameter to 0.00129 in about 24 hours
Anatomy
Oct 12, 2026

Anatomy goes open-source, turning one-line prompts into interactive machines that explain technical concepts

Developer Ryan has open-sourced Anatomy, an AI drawing skill that currently works with Claude Code, OpenCode, and Codex. The tool lets users describe a technical concept in a single sentence, then generates a machine whose moving parts visually explain how that concept works. The resulting output is interactive, with support for mouse and keyboard input, and can be edited as web code, opened on its own, or embedded into a website. Examples shown so far include an LRU cache represented as a bookshelf that moves books, a mechanical watch that can be wound and disassembled, and a Raptor rocket engine used on SpaceX Starship that can be split into 11 component groups and simulate ignition. Anatomy first studies the underlying principles and parameters, then designs parts, draws graphics, and adds animation. Motion and numerical readouts are driven by the same computational model, so they update together during interaction. The project also includes automated checks for issues such as part overlap and occlusion errors. Ryan has packaged existing works into a component library for developers, and the project is available under the MIT license. The developer said detailed builds still take time: the mechanical watch took nearly two hours, while the Raptor engine took more than 10 hours. Its 3D feature remains in testing.

630
Anatomy goes open-source, turning one-line prompts into interactive machines that explain technical concepts
OpenAI
Oct 12, 2026

OpenAI raises GPT-6.1 Sol default context to 500,000 tokens

OpenAI product and platform lead Thibault Sottiaux said the company’s Day 7 update in its 28-day product sprint increases the default context window for GPT-6.1 Sol to 500,000 tokens in Codex and ChatGPT Work, up from 272,000 tokens. Based on the figures provided, that is roughly an 80% increase. Sottiaux said the new default was specifically tested. A larger window cuts down on how often the model has to automatically compress context during long tasks, allowing it to retain more prior conversation and task details. Users can still adjust the setting manually, either higher or lower, with support up to 1 million tokens. OpenAI also said it will no longer charge a long-context surcharge when GPT-6.1 Sol is used within ChatGPT subscription limits. The model had already supported manually setting a larger context window before this change; the latest update affects the default setting rather than adding that capability for the first time.

2500
OpenAI raises GPT-6.1 Sol default context to 500,000 tokens
OpenAI
Oct 12, 2026

NYU professor warns researchers not to put original ideas into OpenAI tools

New York University mathematics professor Tristan Buckmaster is warning researchers against entering original ideas into OpenAI’s AI tools after raising concerns about his own unpublished work. According to the report cited by Techub, Buckmaster suspected that unpublished Navier-Stokes equation research he had been conducting with Anthropic mathematician Levent Alpöge was rapidly reproduced after OpenAI became aware of his name. The two had spent a year working on the fluid dynamics problem, which carries a $1 million prize, and had stored a draft of their research on OpenAI servers. Buckmaster said OpenAI’s AI agent reached the point they had achieved within a day after the company knew his identity, which he described as suspicious timing. An OpenAI spokesperson rejected that claim, saying it could clearly deny that Buckmaster’s recent Codex prompts had any impact on the system. The report also referenced an earlier OpenAI dataset containing 4,000 unsolved problems, which was said to have derailed some young mathematicians’ research projects.

710
NYU professor warns researchers not to put original ideas into OpenAI tools
OpenAI
Oct 11, 2026

OpenAI asks Pro users to choose between a 5-hour cap and doubled weekly limits

OpenAI is polling Pro users on possible usage-limit changes for Codex and ChatGPT Work, according to BlockBeats. The poll was launched by Thibault Sottiaux, the company’s head of product and platform, and presents two options: bring back a usage cap every five hours in exchange for doubling the total weekly allowance, or keep the current setup with no five-hour cap and leave weekly limits unchanged. The proposal stands out because Sottiaux had said on Sept. 29 that the five-hour restriction would not be reintroduced. It has now resurfaced as one of the options under consideration, though only as part of a trade-off tied to higher weekly quotas. No formal rule change has been announced. The current move is limited to gathering user feedback, and OpenAI has not said whether it will revise quota rules for either Codex or ChatGPT Work. The poll reflects an active discussion around how Pro-tier access should be structured, but for now the company has not confirmed any final decision.

780
OpenAI asks Pro users to choose between a 5-hour cap and doubled weekly limits
Total 179 items
Go topage