News 2026-08-14

AI Model & Benchmark Watch — August 14, 2026

Gemini 3.5 Pro missed its third deadline and Google shipped a stopgap Flash model instead, OpenAI is sitting on Astra after finding it might have crossed into "critical" cyber-risk territory, and…

AI Model & Benchmark Watch — August 14, 2026

Gemini 3.5 Pro missed its third deadline and Google shipped a stopgap Flash model instead, OpenAI is sitting on Astra after finding it might have crossed into "critical" cyber-risk territory, and Zhipu's GLM-5.3 stumbled into unplanned exploit-chain reasoning during safety testing — finding 1,097 real vulnerabilities — and delayed its own open-weighting because of it.

Overview

The closed frontier had a rough week for anyone waiting on a headline model, and a genuinely strange one for anyone watching safety disclosures. Google's Gemini 3.5 Pro is now reportedly facing a possible pretraining restart and researcher departures, not just a scheduling slip, and Google shipped Gemini 3.7 Flash on August 13 as a placeholder rather than the model people were waiting for. OpenAI disclosed on August 7 that it "cannot rule out" Astra has reached its "critical" tier for autonomous cyber capability — the first OpenAI model to approach that threshold — and has slowed its development rather than ship it. xAI's Grok 4.6, which launched a week ago with no benchmarks attached, finally got scored by Artificial Analysis on August 12: an Intelligence Index of 61, essentially level with GPT-5.6 Sol. And Zhipu's GLM-5.3, which launched today, turned up unplanned exploit-development capability during its own safety testing, which is also why — for the first time — Zhipu isn't shipping the weights same-day.

New & Updated Models (this week)

Commercial / closed

Gemini 3.7 Flash — Google DeepMind (August 13, 2026)

Who / License: Google DeepMind; closed, hosted via AI Studio, Vertex AI, Antigravity, and GitHub Copilot.

What's notable: Google's own model card calls it "a refinement of 3.6 Flash with algorithmic improvements... not a new pretraining run" — a stopgap, not the frontier update people are waiting for. 1M-token context, up to 64K output, text/image/video/audio input. Artificial Analysis measured its Intelligence Index at 56 on the high-reasoning setting (+4 over 3.6 Flash) and its DeepSWE score jumping from 49.0% to 65.3%. Priced at $0.75/$3.75 per Mtok in/out through the end of 2026, roughly half of 3.6 Flash's rate, rising to $1.50/$7.50 on January 1, 2027.

Source: MarkTechPost — Google AI Just Released Gemini 3.7 Flash · Artificial Analysis — Gemini 3.7 and the time until frontier · Google DeepMind — Gemini 3.7 Flash model card · Bloomberg — Google debuts new Gemini Flash while top AI model still delayed

Gemini 3.5 Pro — still not released

What's notable: Missed its third rumored date (this time August 12). Forbes reports persistent coding/reliability issues, senior researcher departures, and the possibility of a full pretraining restart over a capability ceiling Google is calling structural. No specs, no benchmarks, no ship date.

Source: Forbes — Gemini 3.5 Pro Delay Continues

Astra — OpenAI (still unreleased; safety disclosures August 7 and 11, 2026)

Who / License: OpenAI; unreleased, development slowed.

What's notable: On August 7, OpenAI said it "cannot rule out" that Astra has reached the "critical" tier on its own cybersecurity risk scale — autonomous zero-day discovery and exploitation — a level none of its shipped models, including GPT-5.6 Sol, has hit. In response, OpenAI is running Astra in isolated environments with restricted network and tool access, hardening weight protection, and testing it with government agencies before any release. On August 11, OpenAI added "universal monitoring" — chain-of-thought monitors that can halt the model if it starts down a risky path. No release date, model card, or pricing exists yet.

Source: TechCrunch — OpenAI says it slowed Astra model development over security concerns · The Hacker News — OpenAI's next AI model Astra shows cyber risk · Forbes — OpenAI pauses Astra after it nears first-ever critical cyber risk

GPT-5.6-Cyber — OpenAI (August 10, 2026)

Who / License: OpenAI; closed, tightly gated (identity verification, legal attestations, and — starting September 1 — a hardware security key).

What's notable: Part of a restructured "Daybreak" program: Daybreak Blue strips security-prompt safeguards from existing models like GPT-5.6 Sol for vetted defenders, while Daybreak Red gates this new model for vulnerability research. OpenAI's own numbers show a 95.0% "advanced cybersecurity completion rate" versus 1.5% for a safeguarded GPT-5.6 Sol — but that measures how often the model attempts the task rather than how well, and OpenAI's own ExploitBench data shows GPT-5.6-Cyber actually producing lower-quality, less token-efficient results than plain Sol. OpenAI credits it with finding two chained zero-days in Chrome's V8 engine and hundreds of privilege-escalation bugs in an unnamed OS kernel. No public pricing yet.

Source: VentureBeat — OpenAI launches GPT-5.6-Cyber · eesel AI — GPT-5.6-Cyber · datanorth.ai — OpenAI launches GPT-5.6-Cyber

Grok 4.6 — xAI (launched August 7; full benchmarks published August 12, 2026)

Who / License: xAI; closed.

What's notable: Last week's edition covered the launch with no scores attached. Artificial Analysis's August 12 benchmark run puts its Intelligence Index at 61 (+5 over Grok 4.5), a GDPval-AA v2 Elo of 1753 (behind only Claude Opus 5), and 88.4% on Terminal-Bench v2.1 — while using roughly half the turns and a quarter of the input tokens of Claude Opus 5 on long-horizon tasks. It's still the same 1.5T-parameter "V9" base as Grok 4.5; the gain is entirely post-training. Pricing stays at $2/$6 per Mtok below 200K context, doubling above that. A larger 2.1T-parameter Grok 4.7 is reportedly still weeks out.

Source: Artificial Analysis — Grok 4.6 returns xAI to the intelligence frontier · testingcatalog — xAI releases Grok 4.6 for long-running agent work · basenor — xAI launches Grok 4.6

MAI-Code-1.1-Flash — Microsoft (August 11, 2026)

Who / License: Microsoft; closed, via GitHub Copilot and Microsoft Foundry.

What's notable: A small-tier coding model update adding native vision to what was text-only in MAI-Code-1-Flash, alongside gains in coding quality and tool use. Listed 73% cheaper than its predecessor; billed at a 0.25x premium-request multiplier for annual Copilot subscribers.

Source: GitHub Changelog — MAI-Code-1.1-Flash available in GitHub Copilot

Anthropic shipped no new Claude model this week — only a safety-guardrail update to Fable 5's biology safeguards (August 7) and unconfirmed rumors of a "Fable 5.1" that Anthropic hasn't acknowledged.

Open-weight

GLM-5.3 — Zhipu / Z.ai (August 14, 2026)

Who / License: Zhipu / Z.ai; launched closed via the GLM Coding Plan and ZCode. Open weights promised in roughly two weeks, pending a safety review — a break from GLM-5.2's same-day-open pattern.

What's notable: Post-trained on the same 743B-parameter base as GLM-5.2, with every gain coming from post-training rather than a bigger model. Zhipu reports Terminal-Bench 3.0 jumping from 4.6 to 28.3, DeepSWE v1.1 from 46.2% to 66.9% (still behind GPT-5.6 Sol's reported 72.7%), and CyberGym from 77.2% to 84.5%. The reason for the delayed weight release: safety testing turned up exploit-chain reasoning the team says it didn't intend to train for, and the model reportedly found 1,097 critical vulnerabilities across Linux, WebKit, and FreeBSD during that testing. GLM-5.5, the rumored 1T+-parameter successor people had been expecting this month, remains unconfirmed — Zhipu shipped 5.3 instead.

Source: Tech Times — GLM-5.3 post-training produced exploit chains Z.ai never planned, finds 1,097 critical bugs · BigGo Finance — GLM-5.3 · emergent.sh — GLM-5.3 officially launched

Muse Glimmer — Meta (August 10, 2026)

Who / License: Meta Superintelligence Labs; open, Apache 2.0, ungated on Hugging Face.

What's notable: A 30B dense model distilled from Muse Spark 1.2 with its own perception encoder, compressed to under 20GB at roughly 4-bit precision — small enough to run on a single consumer GPU or a Mac. 128K context, multimodal, aimed at local agentic work rather than benchmark-chasing. Meta says Muse Spark 1.2 itself will eventually be open-sourced too, but that hasn't happened yet.

Source: Bloomberg — Meta releases Muse Glimmer, an AI model people can run on their laptop · Hugging Face — Muse Glimmer · SiliconANGLE — Meta releases open-source Muse Glimmer

Qwen3.8-Max open weights — Alibaba (August 12, 2026)

Who / License: Alibaba; new custom license (qwen3.8-max), reportedly including a revenue-share requirement for large commercial deployments — a departure from Qwen's usual Apache 2.0.

What's notable: Delivers on the date Alibaba promised last week, but this checkpoint (released as Qwen3.8-2.4T-A95B: 2.4T total params, 95B active) is text-only and doesn't carry the 1M-token context or vision the hosted API offers — it's a materially smaller release than the closed product that went GA on August 3. NVIDIA reported over 4,000 tokens/sec per GPU on GB300 NVL72 hardware for the FP8 checkpoint.

Source: llm-stats.com — Qwen3.8-Max open weights · MindStudio — Qwen3.8-2.4T-A95B release · Neomanex — Qwen3.8-Max open weights

DeepSeek V4-Pro-0813 — DeepSeek (GA August 12–13, 2026)

Who / License: DeepSeek; MIT-licensed lineage, but the August build hasn't actually reached Hugging Face yet — the public repo still shows April's preview weights, so this is API-only for now.

What's notable: 1.6T total params, 49B active, 1M context, up to 384K output, with a new speculative-decoding scheme DeepSeek calls DSpark. DeepSeek's own numbers claim 80.6% on SWE-bench Verified and 90.1% on GPQA Diamond; an independent run on vals.ai put SWE-bench Verified at 96.40% instead — see the movement note below, because that gap is large enough to flag rather than average away. Pricing is $0.435/$0.003625 (miss/hit) input and $0.87 output per Mtok, rising to peak/off-peak tiers on August 16.

Source: Unite.AI — DeepSeek ships V4-Pro as its flagship model leaves preview · Tech Times — DeepSeek V4-Pro-0813 goes GA, benchmark claims await independent proof · vals.ai — SWE-bench leaderboard

Head-to-Head — Current Frontier

State of play as of August 14, 2026. AA Index = Artificial Analysis Intelligence Index, pulled from Artificial Analysis's own model pages/articles where marked (primary); other cells sourced from vals.ai's independently-run evals where marked. Figures marked * are self-reported by the vendor. Figures carried from last week's edition (not independently re-checked this week) are marked ‡. "—" means not publicly confirmed or not evaluated on that benchmark.

Model	Org	Open?	AA Index	GPQA Diamond	SWE-bench (Verified/Pro)	Terminal-Bench 2.1	Context	$ / Mtok in/out
Claude Opus 5	Anthropic	No	63.0 (AA, primary)	84.1%‡	97.00% (Verified, vals.ai)	84.64% (vals.ai)	1M	$5 / $25
Grok 4.6	xAI	No	61.0 (AA, primary)	—	—	88.4% (AA)	500K	$2 / $6 (below 200K)
Claude Fable 5	Anthropic	No	62.1‡	—	80.0%‡ (Pro)	88.0%‡	1M	$10 / $50
DeepSeek V4-Pro-0813	DeepSeek	Yes (MIT lineage; Aug weights not yet published)	53‡	90.1%*	96.40% (Verified, vals.ai) vs. 80.6%* self-reported	—	1M	$0.44 / $0.87
GPT-5.6 Sol	OpenAI	No	58.9‡	94.1%‡	64.6%‡ (Pro)	85.77% (vals.ai)	1M+	$5 / $30
Kimi K3	Moonshot AI	Yes (custom license)	57.1‡	93.5%‡	93.40% (Verified, vals.ai)	—	1M	$3 / $15
Claude Opus 4.8	Anthropic	No	55.7‡	—	69.2%‡ (Pro)	~85.0%‡	1M	$5 / $25
Qwen3.8-Max (API)	Alibaba	No (open weights Aug 12 are a stripped-down separate checkpoint)	56.0‡	92.6%*‡	67.7%*‡ (Pro)	86.6%*‡	1M	$2 / $6
GPT-5.6 Terra	OpenAI	No	55.0‡	—	63.4%‡	87.4%‡	1M+	$2 / $12
GLM-5.2	Zhipu / Z.ai	Yes (MIT)	51.1‡	91.2%‡	62.1%‡	82.7%‡	1M	$1.40 / $4.40
Gemini 3.1 Pro Preview	Google DeepMind	No	—	95.45% (vals.ai, GPQA leader)	—	—	—	—

Reading the table: Claude Opus 5 leads on both AA Index and the two independently-run evals (SWE-bench Verified, Terminal-Bench), and holds that lead by a wide enough margin that this week's news didn't touch it. Grok 4.6's first real scorecard lands it at #2 on AA Index, ahead of Claude Fable 5 by Artificial Analysis's number, though Fable 5's own carried-over SWE-bench Pro figure (80.0%) is still the best coding score in this table where a like-for-like comparison exists — Pro and Verified aren't the same test, so don't read Grok as "beating" Fable 5 on coding from this alone. The real oddity is Gemini: the model actually leading GPQA Diamond right now, at 95.45%, is Gemini 3.1 Pro Preview — a model from February — because Gemini 3.5 Pro still hasn't shipped. And DeepSeek V4-Pro-0813's SWE-bench Verified score depends entirely on which source you trust: DeepSeek says 80.6%, vals.ai's own run says 96.40%. That's too big a gap to average, so it's presented as-is rather than picked for you.

Benchmark & Leaderboard Movement

Grok 4.6 finally has a scorecard. A week after launching with nothing attached, it landed at AA Index 61 — ahead of GPT-5.6 Sol (58.9) and last week's Claude Fable 5 figure (59.9), though behind Claude Opus 5. A secondary tracker this week put Fable 5 higher, at 62.1, which would put it back ahead of Grok 4.6 — not confirmed against Artificial Analysis's own page, so treat that ordering as unsettled.
DeepSeek V4-Pro-0813's SWE-bench Verified result doesn't match DeepSeek's own claim. vals.ai's independent run scored it 96.40%, second only to Claude Opus 5; DeepSeek's own release materials say 80.6%. One of these numbers is wrong, or the two evals aren't measuring the same thing — worth watching before anyone treats DeepSeek as the new #2 in coding.
Qwen3.8-Max's open weights are a downgrade from its own API, shipping August 12 as promised but stripped of vision and the 1M-token context, under a new revenue-share license instead of Qwen's usual Apache 2.0 — the opposite direction from most open-weight releases this year.
GLM-5.3 broke Zhipu's same-day-open pattern. Every prior GLM release has shipped open weights alongside the announcement; this is the first held back, specifically because safety testing surfaced exploit-chain capability the team says it didn't train for.
Gemini 3.5 Pro missed a third deadline, and the reporting behind it has shifted from "still polishing" to "possible pretraining restart" — a materially worse signal than a simple slip.
GLM-5.5 remains unconfirmed. The rumored 1T+-parameter successor that some trackers expected this month didn't ship; Zhipu shipped GLM-5.3 instead.

Analysis

For agentic coding, Claude Opus 5 still leads on the only two independently-run evals in this table (SWE-bench Verified, Terminal-Bench), and nothing this week changed that — treat DeepSeek's 96.40% SWE-bench Verified claim as unconfirmed until the discrepancy with its own 80.6% figure gets resolved. For reasoning, GPQA Diamond is close to meaningless as a differentiator right now: the model on top of it is a six-month-old Gemini preview, sitting there only because Google hasn't shipped a real successor. For cheap-and-fast, DeepSeek V4-Pro-0813 ($0.44/$0.87) and Grok 4.6 ($2/$6 under 200K context) both look like good value this week, with the caveat that DeepSeek's own weights for this build aren't even public yet — it's API-only. For open-weight self-hosting, Kimi K3 remains the safest verified pick; Qwen3.8-Max's new open checkpoint trades away exactly the features (context, vision) that made the API version interesting, and GLM-5.3's weights aren't out at all yet.

The bigger story this week isn't the leaderboard, it's what didn't ship. Two of the three closed labs expected to move the frontier — Google and OpenAI — both held back their flagship models, one over apparent technical trouble and one over an actual safety threshold. Meanwhile the open-weight side kept moving, but with real strings attached: Alibaba open-weighted a weaker Qwen than its own API, and Zhipu delayed its usual same-day release after its own model found something concerning in its own testing. "Open" and "closed" aren't converging so much as both getting more cautious, for different reasons.

Sources

Open raw folder ↗ News/AI-News/AI_Model_Benchmark_Watch/August_2026/model-benchmark-watch_2026-08-14

Metadata

Overview

Title	AI Model & Benchmark Watch — August 14, 2026
Category	News
Date	2026-08-14
Slug	model-benchmark-watch-2026-08-14
Description	Gemini 3.5 Pro missed its third deadline and Google shipped a stopgap Flash model instead, OpenAI is sitting on Astra after finding it might have crossed into "critical" cyber-risk territory, and…
Tags	model-benchmark-watch
Source	News/AI-News/AI_Model_Benchmark_Watch/August_2026/model-benchmark-watch_2026-08-14