OpenAI cuts GPT-5.6 Sol developer pricing again
Reuters reported OpenAI cut GPT-5.6 Sol's API pricing by more than 20% on August 21 — the second markdown to Sol's rate this month, after the roughly 50% cut noted in this digest on August 18. OpenAI's own pricing page now lists Sol at $4.00 input / $20.00 output per million tokens, with the promotional rate extended through at least November 21, 2026. It's the latest move in OpenAI's month-long push to undercut rivals on price across the whole Sol/Terra/Luna line, which has already driven downstream moves like Replit's token-free, Luna-powered "Free Mode."
Sources & depth
DeepSeek-v4-flash-vision-exp
DeepSeek quietly shipped an experimental vision endpoint, deepseek-v4-flash-vision-exp, documented in its API guides. It reads JPEG/PNG/GIF/WebP images — up to 600 per request, each up to 32MB inline/by-URL or 64MB via the new Files API — to describe pictures, read screenshot text, and analyze charts, with each image capped at 384 tokens (~800×800px after resizing). It's the first vision-capable member of the V4-flash line since its August 1 debut; the guide doesn't list pricing and open weights still haven't shipped.
Sources & depth
Google DeepMind partners with EVE Online's studio on game-playing AI agents
Google DeepMind announced a research partnership with Fenris Creations, the studio behind EVE Online, to run SIMA 2 — DeepMind's Gemini-powered game-playing agent — inside EVE Online and its spinoffs EVE Vanguard and EVE Frontier. SIMA 2 plays by watching the screen the way a human would and following natural-language instructions, without hooking into game code; the partnership targets continual learning, long-horizon planning, and multi-agent behavior in a persistent, player-driven universe — a tougher testbed than the shorter single-player games (Atari, Goat Simulator, No Man's Sky) SIMA has tackled so far.
Sources & depth
sglang v0.5.18: faster startup, cheaper allreduce, seven new models
SGLang's v0.5.18 release — 710 PRs from 212 contributors — adds serving support for seven new models, including the multimodal Muse Glimmer and video-generation models SANA-Video, LTX-2.5, and Cosmos3 variants. It also ships concrete efficiency wins: checkpoint pages now stream in during CUDA-graph capture for 8.6–11.7% faster startup on Qwen3-32B/H100, a redesigned tensor-parallel LM-head cuts DeepSeek-V4-Pro decode latency from 320μs to 169μs on B200, and FlashInfer workspace reuse adds up to 6.9% at small batch sizes for DeepSeek-V4-Flash on Blackwell. All compiled-kernel caches now share one SGLANG_CACHE_DIR, so upgraders need a one-time recompile.
Sources & depth
- github.comv0.5.182026-08-22 05:39 IST
Ollama adds native model management inside Claude Desktop
Ollama's v0.33.0 release candidates (rc0 through rc2, August 22–23) add a menu-bar toggle to turn individual local Ollama models on or off for use inside Claude Desktop, a picker to choose from available Ollama models directly within Claude (cloud models require sign-in), and a new Apps view for managing integrations with copyable commands. The same cycle fixes prefill cancellation to preserve KV-cache restore points instead of reprocessing, and removes token-countdown system messages that were disrupting cache efficiency.
Sources & depth