Latest AI News

The accessible source details point to this update: OpenAI Cuts GPT-4.5 Sol API and Credit Pricing by Over 20% for Three Months. Because the full article could not be reliably extracted or rewritten, this TLDR stays conservative and is based on headline-level information plus limited source context from OpenAI on X.

xAI has expanded access to Grok Bot, its AI agent platform launched in beta on August 11, 2026, to include SuperGrok Plus, SuperGrok Heavy, Cursor Pro+, Cursor Ultra, and all Cursor Teams plans. Grok Bot functions as a digital teammate that can handle tasks autonomously across apps and inboxes, with browser and terminal access via a cloud computer. Users can run multiple Bots in parallel, set up routines, and message them via a text-thread interface on mobile or desktop.

Midjourney has shipped a major bug-fix update to alpha.midjourney.com, two weeks after launching a rough draft of significant platform changes. The update addresses dozens of issues across folders, upscale, gallery, and performance. Key fixes include collapsible folder groups in the sidebar, restored Upscale, Zoom, and Vary buttons for older V8.2 images, settings that no longer reset unexpectedly, reduced lag in prompt input and gallery scrolling, and corrected pill behavior. Moodboard and profile copy functions are also restored. The team warns new bugs may have been introduced.

Anthropic is expanding Claude Mythos 5 for cyber defense, making it available in Claude Security for Enterprise customers to scan codebases for vulnerabilities and suggest patches. The company is also launching the Defender Advantage Fund, providing $35 million in credits to organizations securing open-source software across three focus areas: patching live vulnerabilities, automating scanning, and pursuing ambitious security approaches. Anthropic is also expanding its Cyber Verification Program to include Mythos-class model access for vetted defenders.

Runway has launched Ruby, a new AI model that converts standard dynamic range (SDR) video to bit HDR, outputting in ProRes and EXR sequence formats. Ruby is compatible with any existing uploaded video or previously generated Runway output, supporting clips up to 30 seconds long. The tool is now available to get started on Runway's platform, offering filmmakers and creators a straightforward path to high-dynamic-range video without reshooting footage.

NVIDIA's Agentic Variation Operators (AVO) research project achieved a perfect 100% score on the ARC-AGI-3 benchmark, completing all 183 levels across 25 environments while using 12% fewer actions than competing system VISTA. AVO elevated Claude Opus 5 from a 30% standalone baseline to 100% by wrapping it in a general-purpose agent architecture featuring persistent memory, a supervisor module, and tool use. The same system previously outperformed FlashAttention-4 by 10.5% on NVIDIA DGX B200 hardware after seven days of autonomous GPU-kernel optimization.

DeepSeek has launched DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model, on its API platform. The model matches DeepSeek-V4-Flash on text capabilities including agents, reasoning, and world knowledge, while making a major leap on multimodal agent benchmarks, approaching Opus-4.8 performance. Images are billed at up to 384 tokens each at V4-Flash pricing. A new Files API lets users upload images once and reuse them via file_id. DeepSeek Harness 0.1.1 was also released with out-of-the-box support for the new model.

OpenCode has launched Ox Alpha, a stealth AI model available free for one week. The model features a 1 million token context window, multimodal capabilities, and zero data retention. OpenCode is offering generous rate limits with near-unlimited usage, backed by infrastructure capable of handling 100 trillion tokens per day. The free access period is intended to stress-test capacity and see how users engage with the model at scale.

OpenAI has added transparent background generation to its GPT-Image-2 API, now available in preview. The feature lets developers create reusable image assets that can be placed on any background, making it useful for product imagery, graphic design, website mockups, and marketing campaigns. OpenAI hinted that more transparency-related features are coming soon. The update expands GPT-Image-2's practical utility for commercial and creative workflows requiring flexible, compositable visual assets.

Anthropic has launched three production-ready tools on the Claude Platform: Computer Use, the Skills API, and the Files API. Computer Use now includes a browser use tool that reads page structure to target web elements more reliably than pixel-based clicking, and supports multi-action turns for faster task completion. The Skills API lets developers upload versioned instruction sets that run in Claude's sandbox. The Files API offers 1 TB storage per organization and 5x higher rate limits. Both APIs are also available via Microsoft Foundry.

Apple Music will add "Made With AI" labels to AI-generated tracks later this year, requiring record labels and distributors to tag songs where a material portion of content was AI-created. The move follows Apple Music's March introduction of AI Transparency Tagging. Apple Music VP Oliver Schusser previously noted over a third of tracks uploaded to the platform are 100% AI-generated. The announcement follows similar moves by Spotify and Tidal, and aligns with advocacy from the RIAA, IFPI, the Grammys, and SAG-AFTRA for industry-wide AI labeling.

Google has launched a Preferred Sources button that publishers can embed on their websites, letting readers designate them as favorite sources to be highlighted more often across Google Search, Discover, and Google News. The feature expands on a May rollout tied to AI Mode and AI Overviews. Google says people are twice as likely to click through to a preferred source when available, and over 345,000 unique sources have already been selected. Google is also adding natural language Discover feed customization and audio briefing personalization for Android users.

Sakana AI has upgraded its free Sakana Translate service to run on the new Sakana Namazu model, which emphasizes deep understanding of Japanese language, culture, and context. The service continues to offer bidirectional translation between Japanese, English, and Chinese, along with proofreading and Q&A features. In head-to-head evaluations using Sakana AI's TransEvalnia tool across 160 Japanese-to-English tasks, Sakana Translate outperformed competing models over 50% of the time. Future plans include file translation for PDFs and Office documents, plus enterprise features like API access and SSO.

OpenAI has launched an Apple Messages plugin for ChatGPT on Mac, allowing users to search messages, catch up on conversations, and draft or send replies using AI. The plugin is now available in ChatGPT Work and Codex on desktop and can be installed via Plugins > Public. Example use cases include checking a calendar to suggest meeting times, finding follow-ups from yesterday, identifying birthdays in messages, and flagging potential spam for deletion.

Adobe Firefly's audio tools are now generally available, bringing Generate Music, Generate Speech, and Generate Sound Effects into its all-in-one creative AI studio. Generate Music uses the Firefly Music Model to create universally licensed original tracks; Generate Speech, powered by the Firefly Speech Model with optional ElevenLabs integration, produces natural voiceovers; and Generate Sound Effects creates custom sounds matched to content. Adobe also added Gemini Omni Flash to its third-party model lineup and introduced a free tier for Firefly AI Assistant with daily generations.

A Pew Research study found that 35% of webpages published after ChatGPT's November 2022 launch show signs of AI authorship, based on analysis of nearly half a million English-language pages from the Common Crawl archive using Open Pangram's detection technology. In a broader sample including older pages, only 10% showed AI signs. Commercial .com domains had AI authorship rates roughly 10 times higher than .edu or .gov sites, which sat near 1%. The findings arrive shortly after Cloudflare reported bot traffic surpassing human web traffic.

Corporate expense management platform Ramp has launched Router, an AI model routing service that lets users switch between large language models via API. Built on three years of internal use, Router supports models from OpenAI, Anthropic, DeepSeek, xAI, and others. It offers benchmark-based routing strategies, a spend dashboard, and a $26 launch credit. The service is free through end of 2026, U.S.-only, and includes an opt-out data retention policy. Ramp, valued at $44 billion, sees Router as an entry point for new customers.

Bengaluru startup Murf AI has launched Falcon 2, a text-to-speech model priced at $0.01 per generated minute, one-fifth the cost of ElevenLabs. Available from August 20, Falcon 2 ranked higher than OpenAI's Realtime API on Artificial Analysis benchmarks, responds in under 100 milliseconds, and supports over 150 voices across 35 languages. Targeting enterprise clients like call centres, banks, and airlines, it can handle 10,000 concurrent calls. Murf AI, founded in 2020 and backed by Matrix Partners and Elevation Capital, serves over 10 million users.

Meta has launched Pocket, an AI-powered game creation app, to U.S. users after quietly testing it in Brazil last month. Built on Meta's acqui-hire of vibe-coding platform Gizmo, Pocket lets users generate small interactive games called gizmos using AI prompts. These games respond to touch and phone tilt, play sound effects, support music clips, and can use camera roll photos. Games are published to a scrollable feed where others can save, remix, or repost them. Pocket joins Meta's growing suite of standalone AI apps.

Salesforce has launched Slack Code, a new feature that brings AI-powered coding collaboration directly into Slack. Code channels give teams a dedicated multiplayer workspace where developers, PMs, and designers can work alongside coding agents like Anthropic's Claude, Cognition's Devin, GitHub Copilot, ChatGPT, and Vercel's agents. Teams can review code diffs, check live previews, and approve work before it ships, all without leaving Slack. Code channels archive automatically when work is done, preserving searchable history. The feature is available on any Slack plan.

Google Ads is expanding its AI Max feature with new A/B testing and Performance Planner tools for Search campaigns. Rolling out in September, advertisers can now test different budgets and ROI targets across multiple Search campaigns in a single experiment. Brand and location controls can remain active during tests, removing a previous barrier. Performance Planner also gains the ability to preview how bidding or budget changes affect campaign performance, with a one-click option to apply suggested changes directly.

Black Forest Labs has launched FLUX Upscale, a standalone tool and BFL API endpoint that regenerates video from 480p up to native 4K. It natively understands FLUX 3 output and corrects artifacts that general upscalers miss, such as smudged faces and texture gridding on water or grass. Two modes are available: Precise at $0.07 per megapixel per second and Creative at $0.10, with upscale factors of 1.5x, 2x, and 3x targeting roughly 1080p, 2K, and 4K.

Perplexity has launched Brain, a structured Markdown-based memory system for its Computer AI agent that functions as a self-improving knowledge wiki. Brain organizes persistent context across three filesystem layers: a synthesized knowledge wiki, distilled notes, and raw session transcripts. Pages use wikilinks to connect related topics and citation references to verify claims against source sessions. Git-backed versioning supports multiple agents updating memory simultaneously. Internal tests showed remote filesystem access was 400–500 times slower than local file operations, informing Brain's hybrid local-plus-semantic-retrieval architecture.

Meta AI has officially launched a macOS desktop app, bringing a standout dictation feature that works system-wide across any application on the computer. Users can activate it by holding the fn key and speaking, with the tool promising ultra-high accuracy transcription. The feature eliminates the need to switch apps or use separate dictation software, making hands-free text input more seamless for Mac users throughout their entire workflow.

Google is offering eligible U.S. college students one free year of Google AI Pro, valued at $19.99 per month, and international students one free year of Google AI Plus. The deal includes new study tools: a dedicated student hub at gemini.google.com/students, AI-powered study notebooks with custom quizzes and Google Calendar integration, interactive 3D visualizations, and Deep Research integrated into Gemini Live for hands-free conversational research. A bundle with YouTube Premium is also available at up to 70% off.

OpenAI is previewing Private Safety Processing, a new system designed to extend safety monitoring across multiple API interactions while remaining compatible with its Zero Data Retention program. ZDR already ensures OpenAI does not retain customer prompts or responses after processing. The new system uses automated tools to detect misuse patterns across related interactions without exposing content to OpenAI personnel. Customer content stays encrypted with customer-controlled keys. A technical white paper and broader rollout are planned for September 2026.

Replit has launched Free Mode, a new way to use its AI Agent that lets Core subscribers create 30x more than before for just $20 per month. Powered by OpenAI's GPT-5.6 Luna, Free Mode allows everyday tasks like chatting and ideating without consuming credits, with usage limits resetting every five hours. Replit also redesigned its interface for faster outcomes and introduced renamed tiers: Power Mode for cost-optimized builds and Max Mode for complex, high-performance work.

ElevenLabs has launched Eleven v3 Conversational, its most expressive real-time speech model, now generally available via ElevenAgents and ElevenAPI. The model supports 70+ languages including German, Spanish, French, Portuguese, and Hindi, and offers a library of 11,000+ voices. Audio tags give developers fine-grained control over emotion and delivery, enabling directions like laughing, whispering, or sarcasm. Pricing starts at $0.05 per 1,000 characters and decreases with scale.

xAI's Grok 4.6 is now generally available on Amazon Bedrock. The model is described as xAI's latest flagship, designed for long-running agents and interactive and visual work. It features a 500,token context window and configurable reasoning effort levels: low, medium, high, and extra high. Pricing is set at $2 per million input tokens and $6 per million output tokens. Grok 4.6 is accessible to all developers in supported AWS regions through the standard Amazon Bedrock interface.

Meta has launched a dedicated Mac app for its AI chatbot, Meta AI, featuring window screen sharing that lets the assistant provide suggestions, answer questions, or create content based on what it sees on screen. The app also supports system-wide dictation across all apps. Alongside the Mac launch, Meta is introducing Google Workspace integration and new tools for businesses and creators, including the ability to analyze Instagram and Facebook post performance, manage ad campaigns, and generate decks, docs, and spreadsheets.

Generalist AI has released GEN-1.5, a robot foundation model that learns new physical tasks from a single demonstration lasting just 3 to 12 seconds, requiring no gradient updates or fine-tuning. Tested across 10 diverse dexterous tasks including unzipping pouches, opening jars, and retrieving money, GEN-1.5 achieves 59% average success via one-shot in-context prompting and 83% with 10 gradient steps on five minutes of data. The model also supports human-to-robot imitation, zero-shot sim-to-real transfer, and compositional generalization across chained behaviors.

Moderna and Merck released three-year follow-up data from their Phase 2b study showing continued clinical benefit of mRNA-4157, an individualized neoantigen therapy, combined with pembrolizumab in high-risk melanoma patients after complete resection. AI algorithms analyze tumor sequencing data to select up to 34 neoantigens most likely to trigger an immune response, while Moderna's Maestro platform manages manufacturing logistics for each patient batch. Phase 3 trials in melanoma and non-small cell lung cancer are now enrolling globally.

Ornith.ai has launched Ornith-1.5, a family of open-source AI models built around an end-to-end self-improvement loop where the model proposes its own training tasks, generates scaffolds, and produces solution rollouts for reinforcement learning. Available in three sizes—397B MoE, 35B MoE, and 9B Dense—the flagship Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, matching Claude Opus 4.8 while outperforming open-source rivals GLM-5.2 and DeepSeek-V4-Flash-0731. The compact 9B model runs on iPhone and Android.

OpenHistory is a free, open-source Mac app created by Zach Tratar that automatically tracks your entire workday and generates hourly and daily AI summaries. All log data is stored locally, and using Apple Intelligence keeps everything on-device. Users can also connect Claude, ChatGPT, or Kimi for better summaries, with a privacy tradeoff. The app connects to local agents via secure MCP and can generate standup notes on demand. The code is fully open-source under the MIT license.

OpenAI is expanding ChatGPT Ads to 31 European markets including Germany, France, Spain, Italy, and Sweden, marking its largest advertising rollout to date. Six months after a U.S. pilot, the platform now reaches tens of thousands of marketers globally. European advertisers can initially access ChatGPT Ads through OpenAI's Ads Solutions team and agency partners, with self-service access via Ads Manager coming later this summer. Ads appear only to Free and Go plan users, while Plus, Pro, and Enterprise subscribers remain ad-free.

Anthropic tested Claude's ability to design protein binders from scratch across 15 targets, succeeding on 14. Using Claude Mythos Preview and Opus 4.8, hit rates reached 22-35% depending on setup, compared to the industry standard of 10-15%. Some designs outperformed the best previously published results. Separately, Claude Opus 5 analyzed NMR and LC-MS chemistry data, completing in under 25 minutes and matching lab purity results of 96.33% versus its own 96.4%. Anthropic plans to launch a dedicated access program for life scientists soon.

Z.ai has launched the GLM-5.3 API, a new model specifically built for coding, defensive cybersecurity, and long-horizon agentic tasks. Priced identically to its predecessor GLM-5.2, the upgrade is cost-neutral for existing users. GLM-5.3 is accessible through Z.ai's official API as well as partner model gateways, giving developers multiple flexible integration options for incorporating the model into their applications, security tooling, and automated workflows.

Anthropic's Claude AI has gained the ability to send emails directly in Gmail and manage files in Google Drive, expanding its real-world action capabilities. Users can ask Claude to reply to an email thread, and it will automatically draft and send the response on their behalf. Users retain control over when Claude requires approval before taking action. The features are available on all paid Claude plans and can be activated by connecting Gmail or Google Drive through the connectors menu in settings.

Gemini in Chrome is now available to all U.S. Android users, functioning as a browsing assistant that summarizes articles, answers questions about web pages, and connects with Google apps like Calendar and Keep without switching tabs. AI Pro and AI Ultra subscribers also gain access to auto browse, an agentic feature that handles tasks like booking parking, updating recurring orders, and organizing travel. Auto browse asks for confirmation before sensitive actions and is designed to detect prompt injection threats.

OpenAI is launching a $5 million initiative to help democratic oversight bodies keep pace with government AI adoption in national security. The program will provide training, technical support, and OpenAI credits to authorized oversight institutions over the next year. OpenAI will also pilot tools helping reviewers examine records surrounding AI-assisted government decisions, including inputs, outputs, and tool use. Participating institutions retain control of findings, and tools will be model-agnostic where feasible.

Perplexity has launched Computer in Email, allowing users to trigger AI agent tasks directly from their inbox by emailing computer@perplexity.com, forwarding a thread, or cc'ing the address. Computer verifies the sender, reads the full thread and attachments, and runs tasks using the sender's own connectors and permissions. Deliverables like Excel models, PDFs, and decks are returned as attachments in the same thread. The feature follows earlier launches in Slack and Teams and is available to all Computer users today.

OpenAI has temporarily paused frontier reinforcement learning training on its upcoming model Astra after preliminary evidence suggested it may meet the Critical cybersecurity capability threshold under the company's Preparedness Framework. Following the OpenAI-Hugging Face incident, OpenAI also hardened research environments, implemented stronger workload and network isolation, and expanded chain-of-thought monitoring with a 20% compute overhead. The largest planned frontier RL run remains on hold while smaller-scale evaluations assess model behavior and validate alignment safeguards before proceeding.

Tripo has launched P2.0 Preview, a 3D model generator built for games and real-time pipelines. It features native quad topology with support for up to 50,000 triangles or 25,000 quads, delivering smooth edge loops around joints and clean quads on flat surfaces. Best suited for game characters, hard-surface props, mechanical assemblies, weapons, vehicles, and high-volume asset pipelines. Every user receives 2 free generations. Tripo P1.0 is now available to all users, and annual plans are discounted up to 50%.

A Pew Research Center survey conducted June 22-28, 2026, finds 52% of Americans are more concerned than excited about AI in daily life, up from 37% in 2021. For the first time, a majority of adults under 30 (55%) share that concern, a record high for the age group. Meanwhile, 71% of Americans expect AI to reduce jobs over the next 20 years, up from 64% in 2024, with 73% of adults under 30 holding that view, up from 61% two years ago.

Stability AI has launched two beta tools for Stable Audio 3.0: a DAW plugin and an enhanced web app at StableAudio.com. The plugin works as an instrument in Logic Pro and Ableton Live, available as macOS AU and VST3 for Apple Silicon and Intel Macs, syncing to session BPM and supporting up to six-minute generations. The web app adds iterative editing, audio-to-audio variation, multi-track mixing with level, pan, mute, and solo controls, and export. Both use commercially-safe models, and users own their outputs.

Google has agreed to pay $10 million to acquire Spirit Airlines' data to train its AI models, following the discount carrier's bankruptcy and full shutdown in May. The dataset includes emails, internal communications, spreadsheets, bookings, frequent flyer records, and HR information, though it has been stripped of personally identifiable details. AI hiring firm Mercor.io placed the second-highest bid at $7.5 million. A bankruptcy court judge is set to rule on the sale Wednesday.

OpenAI has launched ChatGPT for Teens, a dedicated experience for users aged 13–17 with built-in safety protections and a Study Mode that uses guiding questions and step-by-step scaffolding instead of direct answers. Features include responsible homework reminders, quizzes, learning visualizations, and Study Hours that parents or teens can schedule. Default protections cover self-harm, eating disorders, and explicit content, while parental controls allow quiet hours and safety notifications. OpenAI also announced a partnership with CodeAI to help teens understand how AI works.

OpenAI and CodeAI have announced a partnership to build AI literacy among students, coinciding with the launch of ChatGPT for Teens, a dedicated experience designed with learning at the center that includes built-in protections and parental controls. Key initiatives include the Hour of AI program reaching millions of students, a first-ever high school Builders Challenge with OpenAI mentorship, and a joint advisory council on child development. CodeAI's free year-long AI Foundations course will also benefit from OpenAI expert consultation.

Alibaba has launched HappyShrimp 1.0, an AI music generation model developed by its Alibaba Token Hub business group, now available in beta at happyshrimp.ai. The model generates complete tracks from a single text prompt, producing melody, arrangement, lyrics, and vocals without requiring technical music knowledge. It supports genres including Chinese style, pop, R&B, hip hop, rock, and jazz, and can interpret culturally specific prompts. Alibaba also announced a collaboration with China's Taihe Music Group for artist co-creation and content development.

Google has launched Operation Blue Skies, the world's first state-backed trial to avoid contrails at oceanic airspace scale, partnering with the UK Government, NATS, Imperial College London, University of Cambridge, Met Office, and Contrails.org. The month program targets Shanwick airspace over the North Atlantic, which accounts for roughly 5% of global contrail warming. Around 10,000 flights per year will be monitored, with a small percentage slightly rerouted to avoid contrail-sensitive regions. Google UK is contributing £1.4 million in-kind on a pro-bono basis.

Cursor has launched Origin, a code hosting platform now in early beta for all paid plan users. Origin includes repos, pull requests, code browsing, and GitHub sync. Users can create repos at cursor.com/codebase/their-name, or sync existing GitHub repos, which update in real time. PRs sync bidirectionally with GitHub. Integrations with Vercel, Depot, and Buildkite are already available, enabling preview deployments and CI workflows. Agents can answer questions, make changes, and push branches directly within Origin.

Social Media Examiner's 2026 AI Marketing Industry Report, written by founder Michael Stelzner, finds 73% of marketers now use AI daily, nearly doubling from 37% in 2024. Claude tripled its user base from 18% to 65% and overtook ChatGPT as marketers' most important platform, with 42% naming it their top choice versus 39% for ChatGPT. AI agents remain nascent at 11% adoption, but 44% plan to adopt them. Data privacy concerns rose to 77%, nearly matching accuracy worries at 78%.

A 404 Media investigation tracked a rare book shipment using an Apple AirTag, revealing that Amazon buys books in bulk through commercial marketplaces like Biblio, ships them to a Las Vegas warehouse called VGT3, cuts off their bindings, and scans them for AI training data, destroying the originals in the process. Amazon confirmed the purchases, saying it buys books to improve its products and services, including its Nova family of large language models. Printed books are prized because they contain text not available online and predate AI-generated content.

The Motion Picture Association has reached its first-ever AI copyright agreement with ByteDance, whose Seedance text-to-video and Seedream image tools faced cease-and-desist letters from the MPA in February after allegedly generating unauthorized content featuring SpongeBob SquarePants and Stranger Things visuals. ByteDance has since released Seedream 5.0 Pro and Seedance 2.5 with strengthened IP protections. MPA CEO Charles Rivkin called copyright a cornerstone of film and TV, though specific guardrail details were not disclosed.

OpenAI has agreed to secure approximately 8 gigawatts of capacity at the PORTS-Pike Technology Campus in Pike County, Ohio, partnering with SB Energy, NVIDIA, and the U.S. Department of Energy. The six-year buildout through 2032 is expected to create 35,000 construction jobs and 2,500 long-term operating roles. NVIDIA will invest $1.5 billion in SB Energy and support the initial 4.25 GW phase. OpenAI and SB Energy will together contribute $80 million in community grants, plus $84 million in Codex credits for roughly 844,000 Ohio college students.

Unitree has previewed a new humanoid robot nicknamed Superman that can perform a standing high jump of 2 meters and reach a top speed of 12.66 meters per second, surpassing the standing high jump and running speed records of all humans worldwide. The robot has a leg length of 0.85 meters and has been in development for just over three months. Unitree notes there is significant room for further improvement, with more advances expected in the coming months.

White House AI czar David Sacks publicly attacked Anthropic CEO Dario Amodei, accusing him of regulatory capture and fear-mongering. Sacks argues Amodei's push for a federal pre-deployment AI approval agency, variously called an FDA, FAA, or FINRA for AI, would create bureaucratic bottlenecks that handicap the U.S. against China while benefiting incumbents like Anthropic. Sacks also criticized Amodei's unsupported claim that AI would eliminate 50 percent of entry-level knowledge jobs within five years, and accused Anthropic of orchestrating fear campaigns to shape regulation in its favor.

OpenAI Codex now supports a million-token context window using GPT-5.6 Sol, which has a documented 1,050,token limit. Users can enable it by editing ~/.codex/config.toml to set model to gpt-5.sol, model_context_window to 1,000,000, and model_auto_compact_token_limit to 900,000, leaving headroom before compaction. A single-session CLI flag alternative is also available. The larger context lets Codex retain more code, tool output, and conversation history, though OpenAI notes its defaults were carefully tuned for performance and cost.