August 2026 archive
August logged 14 tracked events, 10 of them major. DeepSeek V4 Pro's 2026-08-13 launch chart claimed a win over Claude Opus 4.8 that mostly turns out to be noise (see the verdict); the real story in its own column is a roughly fifty-point jump on DeepSWE from its prior release. GLM-5.3 released on 2026-08-14 and its API followed on 2026-08-19 at an unchanged price — every benchmark number attached to it was vendor-only at launch, and independent evaluators have since run 9 of the 11 scores this site now tracks for it.
Every AI model release in August 2026
- 2026-08-28Release
Tencent Hy4 preview — 770B open-weight at $0.834/$2.501 per 1M tokens, added to tracker retroactively
Official source →Apache 2.0 weights for a 770B mixture-of-experts (49B active) with a 1M-token context, priced flat at $0.834/$2.501. Added to this site on 2026-09-27, a month late.
Tencent's model card reports six scores on benchmarks this site tracks; none has an independent run yet — Artificial Analysis, tbench.ai, Toolathlon, the DeepSWE board, ARC Prize and LiveBench had no Hy4 row on 2026-09-27.
The predecessor offers some reassurance: Tencent's re-run Hy3 figures sit within about a point of independent runs on HLE, GPQA Diamond and Toolathlon-Verified, and inside Terminal-Bench 2.1's noise band on the fourth.
- 2026-08-26Release
Qwen3.8-Flash-Next ships open-weight — a Qwen4-architecture preview under a Qwen3.8 label
Official source →Don't confuse it with the bare 'Qwen3.8-Flash' (no suffix) — Alibaba's own model card describes that as a separate, not-yet-released hosted SKU with production features this preview lacks. Flash-Next is open-weight only, no hosted API price exists yet, and all 4 tracked benchmark scores are Alibaba's own self-reports — none of 8 independent boards checked (Artificial Analysis, vals.ai, ARC Prize, LiveBench, MathArena, DeepSWE, Toolathlon, Snorkel AI) had added it as of launch day.
- 2026-08-26Release
Ox Alpha revealed as GLM-5.3-Flash — GA, MIT license, $0.15/$0.50 per 1M tokens
Official source →Z.ai confirms it outright: 'we tested GLM-5.3-Flash anonymously as ox-alpha ... to gather user feedback.' That settles product identity, not checkpoint identity — LiveBench's 'ox-alpha-max' row (69.2 overall) is still unrenamed as of today, so this site excludes it rather than crediting it to the GA release, the same caution this site applied to a past DeepSeek V4 Pro checkpoint mismatch.
Real paid pricing landed the same day as the reveal (promotional $0.075/$0.25 through 2026-09-09, then $0.15/$0.50), a day earlier than this tracker guessed on 2026-08-20.
- 2026-08-26Price change
GPT-5.6 Sol price drops to $4.00/$20.00 per 1M tokens (promotional)
Official source →Down from $5.00/$30.00 checked here on 2026-08-20 — OpenAI's own pricing page labels this a promotional rate guaranteed "at least through November 21, 2026," not a permanent cut. Long-context tier (>272K input) is $8.00/$30.00, still 2x input / 1.5x output of the new standard rate.
- 2026-08-25Release
GLM-5.3 weights published on Hugging Face — Z.ai keeps its two-week promise
Official source →Recorded here on 2026-10-06, when this site's final review caught that it was still describing the weights as pending: the full checkpoint (142 safetensors shards, 1.4M downloads in its first month) went up on 2026-08-25, roughly the two weeks Z.ai promised at the August 14 launch, with the safety review tied to the model's emergent exploit-finding capability evidently passed.
The license is a custom "other" license named glm-5.3 rather than the MIT of GLM-5.2 or of the Flash-tier GLM-5.3-Flash — the terms live in the repo's LICENSE file.
Z.ai's launch chart claims (Terminal-Bench 2.1 88.2, ALE-CLI 28.5) stay outside this site's tracked set for the same reasons as before: the tracked Terminal-Bench number is Artificial Analysis's 83.9, and the ALE figure is the 105-task CLI split, not the 1,000-item overall this site tracks.
- 2026-08-20Release
Anonymous "Ox Alpha" stealth model appears on OpenRouter
Official source →Listed by an anonymous third-party provider — OpenRouter isn't the maker, and no lab has claimed it. 1M-token context, free during a promotional preview week (real pricing expected around 2026-08-27).
Zero scores on any benchmark board we track (Artificial Analysis, vals.ai, Arena.ai, Toolathlon, Terminal-Bench) as of 2026-08-22 — rechecked, unchanged.
Independent tokenizer fingerprinting now leans toward Zhipu/Z.ai's GLM-5.3 lineage over the earlier Xiaomi MiMo theory, with a third guess (StepFun) surfacing 2026-08-22 — still nobody official has said anything.
- 2026-08-19Price change
GLM-5.3 API goes live — $1.40/$4.40 per 1M tokens, same price as GLM-5.2
Official source →The pricing is the verifiable part: identical to GLM-5.2, cached input $0.26. The launch chart's numbers (88.2 Terminal-Bench 2.1, 66.9 DeepSWE) are still all vendor-run — every score on our GLM-5.3 page stays Unverified until an independent board reruns it.
- 2026-08-14Release
GLM-5.3 released — subscription-only for now
Official source →Z.ai claims +50% coding over GLM-5.2, but only vendor-harness numbers exist and there's no API pricing yet. Verdict pending independent runs.
- 2026-08-13Release
Gemini 3.7 Flash reaches GA
Official source →Second Flash GA in a month; intro pricing through 2026-12-31.
- 2026-08-13Release
DeepSeek V4 Pro (0813) exits preview
Official source →The steepest single-generation jump we've tracked: DeepSWE went from 12.8 (April preview) to 62.7 in four months. The 'beats Opus 4.8' headline is mostly noise on close inspection — see the verdict — but the generational trajectory itself is real.
- 2026-08-12Release
Grok 4.6 released
Official source →Independent boards largely back the launch claims: rank 3 on GPQA Diamond and rank 4 on SWE-bench Verified (vals.ai), and DeepSWE came in above xAI's own number. Agent-board coverage (tbench/Toolathlon/Snorkel) still pending.
- 2026-08-05Release
Muse Spark 1.2 launches — a coding-focused point release in Meta's closed model line
Official source →Independent scores diverge sharply from Meta's own chart: Terminal-Bench 2.1 spans a 13-point range across three sources (Meta 82.9%, Artificial Analysis 80.15%, vals.ai 69.66%), and DeepSWE shows a smaller +4.3pp vendor-favorable gap (Meta 59.3% vs. the independent board's 55%).
Open weights promised 'coming soon' on 2026-08-10 remain unreleased as of 2026-08-26 — Meta open-weighted a separate, smaller model (Muse Glimmer) that same day instead.
- 2026-08-05Deprecation
Claude Opus 4.1 retired from the API
Official source →Exactly on its announced schedule; recommended replacement is Opus 4.8. Anthropic says it retires models to free capacity for new releases, and has committed to preserving the weights.
- 2026-08-03Release
Qwen3.8-Max launches
Official source →Debuted as the highest-scoring new entry on the independent DeepSWE leaderboard at launch — one of the few claims in this batch that's independently confirmed rather than self-reported.