The AI Brief

OpenAI on August 21 cut GPT-5.6 Sol API pricing from $5 to $4 per million input tokens and from $30 to $20 per million output tokens, an input drop of 20% and an output drop of 33%, as a promotional rate valid through at least November 21, 2026. The cut also applies to credits on ChatGPT Work and Codex; Pro, Plus, and Business subscription prices are unchanged. Cached-input pricing fell from $0.50 to $0.40.

At $4/$20, Sol now costs less than Anthropic's Claude Opus 5 on both input ($5) and output ($25). It is OpenAI's second price event in under a month: on July 30 it cut Luna by 80% and Terra by 20%, leaving Sol untouched. Reuters reported the company cited competition from Anthropic and Chinese AI models as the driver. OpenAI had also introduced a Sol Fast mode at 2.5× standard speed, priced separately at twice the standard token rate.

The timing is notable: it follows Anthropic's September 1 Sonnet 5 price increase and DeepSeek's V4-Flash peak-hour hike to $1.32 per million output tokens, which means the competitive middle of the market is consolidating around a narrower price band than existed three weeks ago.

Cognition AI, maker of the Devin autonomous software engineering agent, is in early talks with investors for a funding round targeting a valuation of more than $40 billion, per Bloomberg, citing people familiar with the matter. The round would raise more than $1 billion. The proposed valuation is a more than 50% increase from the $26 billion at which Cognition raised $1 billion in May 2026, less than three months ago. The company's annualized revenue run rate is approaching $1 billion, roughly double what it was at the time of that earlier round.

At the May raise, CEO Scott Wu confirmed an ARR of $492 million and said enterprise usage of Devin had grown 50% month-over-month for six months running. Bloomberg reported in August that customers include Goldman Sachs, Mercedes-Benz, and several US government agencies. If the round closes at $40 billion, Cognition would surpass the private-market valuation Cursor held before SpaceX's $60 billion acquisition, though it would still trail that deal's implied price.

The proposed multiple is approximately 40× annualized revenue, down from roughly 52× at the May raise, a compression that reflects accelerating revenue rather than cooling investor appetite. Terms and timing remain subject to change.

Z.ai launched GLM-5.3 on August 14, 2026, for API access while withholding downloadable weights. The stated reason was unusual: internal evaluations showed the model had developed coherent exploit-chain reasoning, the ability to plan multi-stage attacks rather than merely spot isolated flaws, that the company says was not an intended training objective. Z.ai said it needed roughly two weeks of safety hardening before releasing the weights, targeting approximately August 28.

On the CyberGym benchmark, GLM-5.3 scored 84.5%, up from 77.2% for its predecessor GLM-5.2, placing it slightly ahead of Anthropic's Claude Mythos 5 (83.8%) and OpenAI's GPT-5.6 Sol (83.6%). On ExploitBench, which measures full exploit construction, not just bug identification, GLM-5.3 scored 54.4%, more than double GLM-5.2's 24.4%, though it still trails Mythos 5's 78.0%. Z.ai paired the launch with a public coordinated vulnerability disclosure ledger at cvd.z.ai, reporting 2,436 findings across 269 open-source projects, with 1,097 classified critical or high severity, spanning Linux, WebKit, and FreeBSD. The aggregate figures are Z.ai's own claims; no third-party verification of the full ledger had been published as of publication.

The model reuses GLM-5.2's 744-billion-parameter mixture-of-experts base, meaning every capability gain reflects post-training alone. The US House committees on Homeland Security and China opened a joint inquiry into Chinese AI models in critical infrastructure in April 2026 naming Zhipu AI specifically. Z.ai also launched OpenVuln, a Hugging Face tool that lets open-source maintainers trigger GLM model scans of their repositories.

Anthropic confirmed in a help-center article updated August 11 that Claude models launched on or after August 2, 2026, embed invisible watermarks in generated text and attach C2PA-signed provenance metadata to supported file types including PNG, SVG, and JPG. The mechanism uses a variant of Google DeepMind's SynthID-Text, which subtly biases token-probability distributions at generation time in a pattern that is statistically detectable at the model level. An Anthropic engineer confirmed on August 12 that a text-detection API is in development, that the model itself is not aware it is being watermarked, and that other labs are implementing similar systems.

The watermarks apply across all Claude surfaces, API, Claude.ai, Claude Code, Claude Cowork, and Claude Tag, worldwide. Anthropic explicitly declined to geofence the obligation to the EU despite the rule originating there. Fines under EU AI Act Article 50 reach €15 million or 3% of global annual turnover. The Digital Omnibus package postponed the machine-readable marking duty only for models already on the market before August 2; Anthropic built marking into new models on the original timeline anyway and is working to retrofit older models during the transition period.

Anthropic states that light editing will not fully remove the watermark; only a complete word-by-word rewrite eliminates it. The company cautions that the watermark cannot distinguish between "Claude wrote this" and "Claude heavily edited this," and that file metadata can disappear during format conversion, screenshots, or re-saving. No user or organization information is encoded in the mark, it identifies Claude's probable involvement, not the identity of the requester.

Disclosure: Claude, which generates this brief, is built by Anthropic.

OpenAI on August 20 published the inaugural post on AI Futures, the blog of its new Strategic Futures team. The team describes its collective goal as answering one question: how should free society be restructured to preserve individual rights and agency while accommodating the emergence of transformative AI. The first post, authored by Dean Ball, frames the team's work explicitly around what the AI safety and policy community calls concentration-of-power risks, scenarios in which AI capability enables a small group to exert unprecedented control over economies or governments.

OpenAI structured the blog so that individual posts carry author views, not organizational positions, a separation it stated explicitly in the inaugural entry. The distinction is consequential: it allows the team to engage politically sensitive governance questions that could complicate regulatory or investor relationships if attributed to OpenAI as an institution. The team is small and not yet publicly staffed beyond Ball's byline on the founding post.

The timing sits between OpenAI's disbanding of its Preparedness team at the end of July, the third dedicated safety structure dissolved in two years, and the expected opening of its public S-1 filing window. The 1,178-person "Pacing the Frontier" letter, signed in part by OpenAI employees on July 28, had called for government tools to implement a verifiable international AI slowdown mechanism; Strategic Futures operates on a domestic US institutional-reform frame rather than an international pacing one.