AI Model & Benchmark Watch — August 21, 2026

Nobody shipped a flagship this week — Gemini 3.5 Pro, Astra, and Grok 4.7 are all still in the oven — so the week's most interesting release is a 27B dense model you can run on a laptop, and Zhipu put an actual date on GLM-5.3's delayed open weights.

Overview

This was the quietest week this beat has covered since it started tracking the current cycle. None of the three delayed flagships — Google's Gemini 3.5 Pro, OpenAI's Astra, xAI's Grok 4.7 — moved from "still not out" to "out," and none of them picked up much new detail beyond confirming they're still not out. The one genuine launch, Alibaba's Qwen3.8-27B on August 14, is a 27.8-billion-parameter dense model built to run on a single consumer GPU rather than to top a leaderboard, and it reportedly hit a million downloads in its first day. Zhipu, meanwhile, gave GLM-5.3's held-back open weights a real date — around August 28 — after last week's edition covered the safety review that delayed them in the first place.

New & Updated Models (this week)

Commercial / closed

Nothing new shipped from a closed lab this week. What moved instead is the delay math on three models already covered:

• Gemini 3.5 Pro remains in limited Vertex AI enterprise preview. No new specifics surfaced beyond what last week's edition covered — the coding-quality problems, the researcher departures, and the possibility of a pretraining restart. Still no ship date.
• Astra had no material update this week. OpenAI's August 7 "cannot rule out critical cyber risk" disclosure, covered last week, still stands as the most recent word on it; Polymarket traders now put the odds of a public launch by August 31 at roughly 18%.
• Grok 4.7 picked up one real data point: as of August 13, Elon Musk said initial training on the rumored 2.1-trillion-parameter model is complete, and xAI is now running a supplemental training pass on SpaceX engineering data. The release window narrowed to "3 to 4 weeks," which points to early September — itself a slip from the late-July estimate. Still no model ID, pricing, or benchmark card from xAI directly; everything traces back to Musk's own posts. Separately, Grok 4.6 landed on Amazon Bedrock this week, widening its distribution without changing the model itself.

Anthropic shipped no new Claude model. The week's Claude news was operational: a roughly 40-minute outage across Claude.ai, Claude Code, and Claude Cowork on August 16, and a new "Concise" output style for Claude Code on August 20. The Fable 5.1 rumor covered in recent weeks is still unconfirmed — no model card, pricing page, or benchmark entry exists for it.

Open-weight

Qwen3.8-27B — Alibaba (August 14, 2026)

Who / License: Alibaba's Tongyi Lab; open, Apache 2.0, published on Hugging Face.

What's notable: A 27.8B-parameter dense multimodal model — small enough to self-host on roughly 16–17GB of RAM or VRAM — with native image and video understanding, a 262K-token native context window extendable to 1M via YaRN, and an adjustable reasoning-effort setting. Alibaba's own numbers put it at 89.2% on GPQA Diamond, 61.7% on SWE-bench Pro, and 73.0 on Terminal-Bench 2.1. It's aimed squarely at local, self-hosted use rather than the frontier leaderboard, and it isn't in the head-to-head table below for that reason — it's a different weight class. It reportedly passed a million downloads within its first 24 hours on Hugging Face.

Source: Local AI Zone — Qwen3.8-27B: A Comprehensive Technical Analysis · Yotta Labs — Qwen 3.8 27B: Specs, Hardware Requirements · StartupHub — Qwen3.8 27B Challenges GPT-5.6 and DeepSeek V4 in Benchmarks

GLM-5.3 open weights — Zhipu / Z.ai (weights expected ~August 28, 2026)

Who / License: Zhipu / Z.ai; API shipped closed on August 14 (covered in last week's edition), open weights now dated to roughly two weeks later.

What's notable: Nothing about the model changed this week, but the timeline firmed up: Z.ai committed to releasing weights around August 28, pending completion of the safety-hardening pass triggered by GLM-5.3's unplanned exploit-chain capability. That review is reportedly the most extensive Z.ai has run on a release to date. Whether August 28 holds is still a promise, not a guarantee — worth checking next week.

Source: MindStudio — When Will GLM 5.3 Open Weights Be Released? · implicator.ai — Z.ai Delays GLM-5.3 Weights After CyberGym Score Tops Mythos · Kingy AI — GLM-5.3's Open-Weight Reality Check

Head-to-Head — Current Frontier

State of play as of August 21, 2026. AA Index = Artificial Analysis Intelligence Index, re-pulled this week directly from Artificial Analysis's own leaderboard (primary). GPQA Diamond and SWE-bench/Terminal-Bench cells cite vals.ai's independently-run evals where marked, or are self-reported by the vendor where marked *. Figures carried from last week's edition without independent re-check this week are marked ‡. "—" means not publicly confirmed or not evaluated on that benchmark.

Reading the table: Claude Opus 5 still leads the field, holding the top AA Index score and both independently-run coding evals in this table for a second straight week — nothing this week challenged that. The bigger news is GLM-5.3 breaking into the AA Index top tier at 60, tied with Kimi K3, purely on post-training gains over the same 744B base as GLM-5.2 (which scored 51.1 last week) — a reminder that post-training alone can move the needle more than a bigger model. GPQA Diamond is still stuck on the same oddity as last week: the actual leader is Gemini 3.1 Pro Preview, a model from February, because Gemini 3.5 Pro still hasn't shipped to replace it. DeepSeek's SWE-bench discrepancy (96.40% independently measured vs. 80.6% self-reported) is unresolved and still worth treating with caution.

Benchmark & Leaderboard Movement

• GLM-5.3 entered the AA Index top tier at 60, up sharply from GLM-5.2's 51.1 — the same 744B base, all post-training. It now sits level with Kimi K3 and just one point behind GPT-5.6 Sol and Grok 4.6.
• Grok 4.7's window slipped again. Musk's own timeline moved from a vague "August" target in late July to "3–4 weeks" as of August 13 — early September at the earliest — even with initial training reportedly complete.
• Astra's status is unchanged from last week, which is itself the news: no new disclosures, no walkback of the "cannot rule out critical cyber risk" finding, and prediction markets have settled around an 18% chance of an August 31 public launch.
• A small open model made the week's real splash. Qwen3.8-27B isn't a frontier contender by AA Index, but a reported million downloads in 24 hours says more about where a lot of real-world usage is heading — cheap, local, good-enough — than any leaderboard entry this week.
• GLM-5.3's open weights now have a target date (~August 28) instead of an open-ended "in about two weeks," first reported last week.

Analysis

For agentic coding, Claude Opus 5 is unchanged at the top for a second week running on both independently-run evals in this table; DeepSeek's 96.40% SWE-bench Verified claim remains unresolved against its own 80.6% figure and shouldn't be read as a real challenge yet. For reasoning, GPQA Diamond is still led by a six-month-old Gemini preview by default, not by merit — that's a statement about Google's release cadence, not about the benchmark. For cheap-and-fast, GLM-5.3's jump to an AA Index of 60 at $1.40/$4.40 per Mtok makes it one of the best value plays in the table right now, though its weights aren't self-hostable until roughly August 28. For local and self-hosted, Qwen3.8-27B is this week's actual pickup: a 27B dense model that runs on a single consumer GPU and still clears 89% on GPQA Diamond.

The larger pattern is a frontier that's paused at the top and still moving underneath. Three closed labs are each sitting on a model they haven't shipped, for three different reasons — coding quality, cyber-risk review, and supplemental training — while the open-weight side keeps closing distance through smaller, cheaper, or purely post-trained releases. Nobody set a new #1 this week. Several labs quietly got closer to the ones that already exist.

Sources

• Forbes — Gemini 3.5 Pro Delay Continues
• eesel AI — Gemini 3.5 Pro: is it out yet?
• Axios — Exclusive: OpenAI slows release of Astra model citing cyber capabilities
• Polymarket — OpenAI's Astra released by…?
• Proactive Investors — Punters bet OpenAI's delayed Astra model will not launch until late 2026
• OrcaRouter — Grok 4.7 Release Date: 3–4 Weeks, and Already Slipping
• kie.ai — What Is Grok 4.7? The 2.1T xAI Model Explained
• Big Hat Group — xAI Weekly: Musk Targets August for Grok 4.6 and 4.7
• Releasebot — xAI Release Notes, August 2026
• Releasebot — Anthropic Release Notes, August 2026
• Releasebot — Claude Code Updates, August 2026
• BleepingComputer — Anthropic confirms Claude is down in major outage
• AIToolsReview — Claude Fable 5.1 Fact Check: What's Actually Confirmed
• CellCog — Fable 5.1: Release Date, Rumors, and What We Actually Know
• Local AI Zone — Qwen3.8-27B: A Comprehensive Technical Analysis
• Yotta Labs — Qwen 3.8 27B: Specs, Hardware Requirements, and How to Run It
• OrcaRouter — Qwen3.8-27B Release Date: Weights Dropped Aug 14, 2026
• StartupHub — Qwen3.8 27B Challenges GPT-5.6 and DeepSeek V4 in Benchmarks
• MindStudio — When Will GLM 5.3 Open Weights Be Released?
• implicator.ai — Z.ai Delays GLM-5.3 Weights After CyberGym Score Tops Mythos
• Kingy AI — GLM-5.3's Open-Weight Reality Check
• VentureBeat — GLM-5.3 hits the API at $1.4/$4.4 per million tokens
• Artificial Analysis — LLM Leaderboard
• vals.ai — SWE-bench leaderboard
• vals.ai — GPQA leaderboard
• vals.ai — Terminal-Bench 2.1 leaderboard
• vals.ai — Kimi K3 model page
• codersera — Kimi K3 Benchmarks: How It Stacks Up vs Fable 5, GPT-5.6 Sol & Opus 4.8
• fenxi.fr — Claude Opus 5 vs GPT-5.6 Sol vs Fable 5 vs Kimi K3 in 2026