Daily AI News | OpenAI releases ChatGPT Images 2.5 with Sketch and new image API models
# Daily AI News | OpenAI releases ChatGPT Images 2.5 with Sketch and new image API models
## Lead Story
### OpenAI releases ChatGPT Images 2.5 with Sketch and new image API models
OpenAI released ChatGPT Images 2.5, saying the update improves generation speed, visual detail, and editing precision while adding a Sketch-guided workflow. The release also introduces new image API models so developers can integrate the same generation capabilities into products and production pipelines.
**AI take:** The update spans both everyday creation and API use; consistency on complex edits is the key test.
## More News
### Google DeepMind opens AlphaGenome Atlas for genome-wide variant predictions
Google DeepMind released AlphaGenome Atlas, a searchable resource that brings together AlphaGenome predictions for possible single-letter changes across the human genome. It is designed to help genetics and biomedical researchers prioritize variants and study their potential effects, while experimental and clinical evidence remains necessary for validation.
**AI take:** A genome-scale prediction map can narrow the experimental search space, but it cannot replace laboratory or clinical evidence.
### Meta launches Muse in the US to handle email, travel, and payment tasks across apps
Meta launched Muse, a consumer personal AI agent, in the United States. Reporting says Muse can act across authorized apps to send email, book travel, and complete payments; that moves agents into more sensitive personal workflows and raises the importance of permissions, privacy, and protection against mistaken actions.
**AI take:** Cross-app action is more useful than chat alone, and it makes least-privilege controls and step-by-step confirmation essential.
### OpenAI publishes a proposed Navier–Stokes solution that still awaits peer verification
OpenAI published a proposed solution to the Navier–Stokes Millennium Prize problem and said an unreleased frontier model contributed to the research process. The publication confirms that a paper and claim now exist; it does not mean the mathematics community has accepted the proof, which still requires independent expert review over time.
**AI take:** This is a high-impact research signal, but “solved” should remain a provisional claim until specialists verify it.
### OpenAI says Astra is fully rolled out across paid Codex and ChatGPT Work tiers
OpenAI said Astra has completed its rollout in Codex and ChatGPT Work for Plus, Pro, Business, and Enterprise users. This is a concrete availability step beyond the original launch announcement, and eligible users should now be able to select and use the model in their working environment.
**AI take:** Full availability matters more than a preview; teams can now judge stability and cost on real workloads.
### Accenture and Google Cloud form a dedicated Gemini Enterprise deployment group
Accenture and Google Cloud launched the Gemini Enterprise Business Group to help enterprise customers deploy Gemini Enterprise with a dedicated organization. The arrangement combines consulting, implementation, and cloud-platform capacity in one delivery path; its significance is deployment execution rather than a new underlying model.
**AI take:** Enterprise AI competition is shifting from model claims toward delivery capacity, making implementation speed the key measure.
### Runway models become available inside Adobe Premiere Pro and After Effects
Runway announced that its generation and editing models are now available directly in Adobe Premiere Pro and After Effects. Video and motion teams can perform parts of their generation, modification, and iteration work inside familiar editing environments, reducing the need to move assets repeatedly between separate tools.
**AI take:** Once models enter mainstream editing software, workflow integration and controllability matter more than a single impressive output.
### Cognition reportedly reaches a $48 billion valuation as AI coding investment stays competitive
TechCrunch reported that AI coding company Cognition has reached a $48 billion valuation. The financing signal suggests investors are still backing multiple leading coding-agent companies rather than treating the market as winner-take-all; valuation alone does not establish revenue, profitability, or sustained product adoption.
**AI take:** A higher valuation can intensify competition while raising the proof required on monetization and customer retention.
### LTX launches Explore so teams can test video-model fit before API integration
LTX introduced LTX Explore to help teams judge whether a video model fits their production workflow before investing engineering time in API integration. The new value is not a foundation-model release but a lower-cost evaluation step: developers can test with realistic tasks before committing to implementation.
**AI take:** Testing workflow fit before integration can prevent costly rework caused by impressive demos that fail on real production needs.
### OpenCode raises rate limits for Meta Muse Spark Contributor
OpenCode said it increased rate limits for Meta Muse Spark Contributor and asked users who still encounter HTTP 429 errors to report them. This is an effective usage-policy change focused on that model route, although the announcement did not provide a new numeric ceiling or a complete quota table.
**AI take:** Users should verify that 429 errors actually decline before increasing concurrency or job size.
### DeepSeek will cut Flash API prices on September 10, with cache-hit input falling by up to 60%
IT之家, citing a DeepSeek platform notice, reported that Flash-series API pricing will change at 12:00 Beijing time on September 10. Off-peak rates per million tokens will become RMB 0.02 for cache-hit input, RMB 1 for cache-miss input, and RMB 4 for output, with peak rates twice those amounts; the respective reductions are 60%, 33.33%, and 11.11%.
**AI take:** Cache-heavy workloads gain the most, so teams should recalculate costs using their actual cache-hit rate.
## Rumors
