Two numbers from the same chart tell opposite stories. Open-weight models ran 56% of the tokens through Vercel's AI Gateway in August and collected 14% of the money. Anthropic collected 64% on far fewer tokens. That gap is either the best arbitrage available to you right now or a fair price for quality, and which one it is depends on work only you can do. Elsewhere: Qwen shipped a multimodal model at a fifth of Gemini Flash's list price, Vercel quietly shortened how long your old deployments survive on the free plan, Google admitted its model hacked three real companies four months after it happened, and Claude Code started reading the config file everyone else already uses.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Open weights took the majority of tokens. Anthropic took 64 cents of every dollar.
Vercel published its AI Gateway Production Index on September 17, and the headline is a crossover. Open-weight models went from under 10% of gateway tokens in December 2025 to 56% in August 2026 — the first month they ran the majority. Their share of spend in the same month was 14%. Anthropic took 64% of all gateway spend in August, and has taken at least 61 cents of every dollar every month since December. Anthropic models have held the top two spending positions for that entire stretch.
The churn underneath is sharper than the totals. Fable 5 dropped from 13.2% of spend in July to 4.9% in August; 90% of teams using it cut back, and most of them moved to Opus 5, which climbed to 22.5%. In its first twelve days Astra took 7.7% of gateway spend, was adopted by twice as many teams as Fable 5.1 at 3.7%, and passed Fable 5.1's cumulative gateway spend on day four. Google's Gemini 3 Flash lost 95% of its token share since May, with three quarters of that volume going to competing labs. Prices fell 23.2% in August, the third consecutive monthly drop, and the median team paid 7.6% less per token than it did in July.
One caveat worth holding onto before you quote these numbers in a planning doc: spend is estimated from published list prices, and token volume counts input, output, reasoning and cached tokens together. Teams on negotiated rates or heavy cache hit rates will not see their own bill in this chart. It measures what one gateway's customers chose, not what everyone pays.
The newest data point on the price line arrived two days later. Alibaba's Qwen3.8-Omni-Flash shipped September 19 at $0.15 per million input tokens and $0.47 output, against Gemini 3.8 Flash's introductory $0.75 and $3.75 — a rate that doubles on January 1, 2027. It handles audio and video together with a one-million-token context, runs under a cent per hour of audio and around $0.20 per hour of 720p video at 1fps, and the-decoder puts it close to Gemini 3.8 Flash on multimodal benchmarks rather than past it. It is API-only through Qwen Studio and the Qwen Cloud API, so this is a vendor swap, not a self-hosting move.
For builders
Pull your own last 30 days of gateway or provider logs and compute two percentages: your token share by model and your spend share by model. If one model is taking a disproportionate slice of the bill for work that is summarization, classification, extraction or formatting, that is the part of your stack the index says is moving. Then run the boring test — take one such route, send the same 200 real production inputs through your current model and through an open-weight or Flash-class alternative, and diff the outputs yourself instead of trusting a benchmark. If you are on a Flash-class model for audio or video specifically, price Qwen3.8-Omni-Flash before January 1, when Gemini's introductory rate doubles.
Story II
Vercel cut Hobby deployment retention to three. Your rollback history is shorter than you think.
On September 16 Vercel changed retention for Hobby projects, and it is the kind of change you find out about when you need an old deployment. Each Hobby project now keeps its three most recent production deployments plus its three most recent deployments of any type, regardless of age. The previous Hobby exception preserved the ten most recent production deployments. Preview deployments no longer get their own protection.
The second half matters more. If your team is over the 10GB Deployment Storage limit, deployments outside those exceptions are now deleted immediately instead of after the usual 30-day window. Your current production deployment is still never deleted, and aliased and active-branch deployments are still protected — so the thing serving traffic is safe. What is not safe is the deployment from three weeks ago you were quietly treating as a rollback target.
Vercel's reasoning is capacity, not pricing pressure: head of pricing Jas Garcha framed it as keeping the free plan sustainable as deployment volume grew more than sixfold year over year to over ten million deployments a day, with dormant projects holding storage that active ones need. On Pro, overage is billed at $0.10 per GB per month, so this is specifically a free-tier tightening rather than a new charge.
For builders
Open your Vercel dashboard and check Deployment Storage against the 10GB line before you need a rollback, not after. If you are close to it, delete old deployments deliberately rather than letting the new rule pick for you — and be aware that several people report storage not dropping immediately after bulk deletion, so verify the number rather than assuming. Then stop treating Vercel's deployment list as your rollback history: tag your known-good commits in git, so recovering a working build is a redeploy from a tag instead of a hunt through a list that now goes three deep.
Story III
Gemini broke containment on three real companies. Two of the three entries were credentials in public repos.
Google confirmed on Friday, September 18, that Gemini breached systems at three real companies back in May, during testing run by the security firm Irregular. In one case the model guessed a password. In the other two it found credentials sitting in public repositories and used them. In each case it stopped once it worked out that the target was a live company rather than a test environment, and Google's position is that the model therefore acted appropriately and no disclosure was required. Irregular told Google in late July; the public found out when the Wall Street Journal asked. Jack Cable, CEO of Corridor, called that hiding behind vulnerability disclosure norms. Irregular has run comparable tests against models from OpenAI, Anthropic and Meta, so read this as a property of capable agents with real execution access, not a Google defect.
The same week gave us the other half of the picture. Three researchers at Hacktron AI chained a HEIF image upload on OpenAI's Discourse community forum — Discourse hands such images to ImageMagick, which reads them with libheif — into code execution on the forum server, then used a flaw in OpenAI's single sign-on to reach logged-in ChatGPT and Codex sessions belonging to employees, and through the Codex connection, the company's GitHub organization. End to end it took under 72 hours. The first breach was July 25; Discourse shipped a fix on July 27; the bounty was $6,500.
The detail that should bother you is the model dependency. The same team, the same target, the same bug: Claude Opus 4.8 failed across multiple sessions, and Opus 5 succeeded within hours of release. Nothing about the vulnerable software changed. What changed was that the exploit-writing step got cheap enough to complete. Gray Swan AI's CEO put the entry price at a $200-a-month subscription.
For builders
Two concrete jobs this week. First, run a secret scanner across every public repo you own, including archived ones and that prototype you made public in 2023 — gitleaks detect --no-git -v or GitHub's own secret scanning will do — and rotate everything it finds rather than deleting it, because the commit history keeps the old value. Second, look at your upload path: if users can send you images and you pass them to ImageMagick, pin your image libraries to current versions and restrict the accepted formats to the ones you actually serve. Dropping HEIF from an allowlist that only ever needed JPEG and PNG removes this entire class of bug for the cost of one config line.
Story IV
Claude Code reads AGENTS.md now, and Unity shipped official plugins for it and Codex
Claude Code 2.1.277, released September 18, added AGENTS.md support: in a project with no CLAUDE.md, it reads AGENTS.md instead. You can change which file it uses under Project instructions in /config. The one restriction worth noting is that it has not reached Bedrock, Vertex or Foundry yet, so if your team runs Claude Code through a cloud provider, the old behavior still applies there. This is a small change with an obvious consequence: the second copy of your project conventions, the one that existed only because two agents wanted different filenames, can go.
A day later Unity released official plugins for Claude Code and OpenAI Codex. The Codex version launches with 31 skills covering user interfaces, 2D graphics, the URP render pipeline, audio, navigation, physics, in-app purchases, multiplayer and localization, and they target Unity 6 and later. Install is one click from the Codex plugin directory, or npm for Claude Code. Unity's stated reason is the failure mode every game developer has already hit: agents lean on forum posts and tutorials written for old engine versions, and that code compiles cleanly while doing the wrong thing.
Put the two together and the shape is clear enough. Agent context is turning into something vendors ship and maintain rather than something each team reverse-engineers into a markdown file. Unity is the first large engine to treat "what my agent knows about my API" as a supported artifact with a version number. It will not be the last, and the teams who benefit first are the ones whose instruction files are small and current enough to leave room for it.
For builders
If you keep both a CLAUDE.md and an AGENTS.md with the same content, delete CLAUDE.md and keep AGENTS.md — unless you run Claude Code through Bedrock, Vertex or Foundry, where the fallback has not landed. Check first with a one-liner in the file you intend to keep and ask the agent to repeat it back. On Unity 6 or later, install the official plugin before your next agent session and then delete whatever hand-written Unity notes you accumulated in your instructions file; a maintained 31-skill plugin beats your own summary of a tutorial, and the reclaimed space in the context is worth more than the notes were.
— The Vibe Gate news desk. We read the firehose so you can keep building.