Ten items today, and the through-line is price. Buyers are trading down, capability is getting cheaper faster than budgets are growing, and the market is repricing what an AI-era software business is worth.
- The harness is now a primitive. OpenAI put the machinery behind Codex into a public-beta API, and the scaffolding your team hand-built became a vendor feature.
- Sixty-eight cents. Ramp's August index puts the effective price of a million tokens down 41% since March, with frontier models losing share to cheaper ones.
- Seven labs, thousands of accounts. Anthropic named Chinese labs that distilled Claude, two of which resold it to their own customers as their own model.
- Open to slowing down. Sam Altman told staff OpenAI would consider pausing frontier development, days after two researchers said the labs are not acting responsibly.
- Ninety percent off. Bending Spoons agreed to buy Miro for $1.36 billion, against a $17.5 billion mark in late 2021.
- The revenue never left. Stripe's index of 72,000 non-AI SaaS businesses shows growth above 30% year on year, through and after the valuation collapse.
- Running out is a feature. An Every editor argues that hitting his model limits is what stops him from building things that should not exist.
- Ninety-three percent of a catalog. Pocket FM doubled its run rate to $500 million with AI producing nearly all new audio content.
- Ask your own data. ChatGPT Work gained a data agent that queries connected company sources and builds shareable dashboards in the chat window.
- Did the ad cause the sale? Google shipped Meridian GeoX to general availability, an open-source way to test incrementality across any platform.
The pattern worth pricing: capability is deflating while the assets built on top of it are being marked to a new, lower reality. Plan your next twelve months on cheaper tokens and stingier buyers.
Developments
Product & Technology
September 10: OpenAI put the Codex harness behind one API call
OpenAI opened its Agents API to all developers in public beta on September 10, exposing the same managed harness that runs Codex and ChatGPT for Work. The design reduces to four objects: an agent definition carrying model, instructions, tools and MCP servers; an optional sandbox environment; a durable session; and the event stream that session returns. It handles context compaction, tool search and subagent coordination, and keeps runs alive for days. Compute can sit in OpenAI's sandbox, your own infrastructure, or a partner's, with Cloudflare, DigitalOcean and Oracle integrations at launch.
So what
The orchestration layer your engineers have been hand-rolling for a year is now a vendor primitive, priced at token cost. That is good news if your advantage is proprietary data, permissions and exception handling, and a problem if your product was the harness itself. Pull your agent roadmap this month and mark every line item that the Agents API now does for free. Redirect that engineering capacity toward the parts a general-purpose API cannot know about your customers.
Confidence: Established fact — OpenAI published the launch, the object model and the pricing, and the beta is open.
Go deeper OpenAI · The Neuron
Operations & Finance
September 9: AI spend per employee fell 10%, and the token got 41% cheaper
Ramp's August AI Index, drawn from card data across roughly 70,000 companies, found median AI spend at the top 1% of adopters fell from about $7,400 to $7,205 per employee. The effective price of a million tokens dropped to $0.68, down 41% from the March peak of $1.15. Frontier models took 45% of token share, down from roughly 53% in early August, as buyers defaulted to Sonnet-class and Terra-class models. Overall adoption still grew, but only 0.4 points, to 56.1% of US businesses. Ramp titled the release "Cracks in the AI thesis."
So what
If you sell AI, your customers' bills are shrinking without them buying less, which means net revenue retention now depends on usage growth outrunning price decay. If you buy AI, this is the number your board will hold your spend against. Separate your own AI line item into price and volume before your next board meeting, and report them apart. A flat bill on rising usage is a win that a single spend figure will hide.
Hype check
Ramp's panel skews toward technology companies and startups, so this is not the US economy. Ramp's own chief economist notes that August spend also falls for ordinary seasonal reasons, as it does in November and December. One month is a data point, not a trend.
Confidence: Established fact — the figures come from Ramp's published index, with the sampling caveat above.
Go deeper Ramp · The Deep View
AI Strategy & Transformation
September 10: Anthropic named the labs distilling Claude, two of which resold it as their own
Anthropic's September threat report covers misuse disrupted between December 2025 and August, and it names seven Chinese labs running distillation through thousands of fraudulent accounts, including Alibaba, DeepSeek, Moonshot and Xiaomi. In some instances Moonshot and DeepSeek served Claude responses to their own paying customers as their own model, then trained on the output. The report also documents a Yemen operation that used Claude Code to build rocket guidance software, a consultant who built Mali's spy agency a system aimed at 25 million phone lines, and five biology cases flagged for possible weapons applications. Anthropic states it does not assert those scientists intended harm.
So what
The resale finding is the commercially load-bearing one. If a cheap model you are evaluating may be a wrapper around a frontier model obtained through fraudulent accounts, your cost advantage is contingent on someone else's enforcement, and your data path is not what the contract says. Add two questions to model procurement: where do the weights come from, and does the vendor warrant that inference runs on the model named in the agreement. Price is a poor reason to accept an unverifiable supply chain.
Confidence: Informed perspective — Anthropic published the evidence, but it is an interested party naming competitors, and the accused labs have not responded.
Go deeper Anthropic · The Rundown AI
AI Strategy & TransformationDeveloping
September 11: Altman told staff OpenAI is open to slowing cutting-edge development
Bloomberg reported that Sam Altman told OpenAI employees the company would consider slowing frontier development on safety grounds. It follows a week in which Anthropic researcher Jacob Coxon resigned saying both labs are "gambling with our lives," Anthropic alignment lead Evan Hubinger put extinction risk above 10% within a decade, and OpenAI Foundation board member Paul Christiano said the industry is not on track to reduce loss-of-control risk to an acceptable level. We covered the resignation on Thursday; the new element is a chief executive entertaining a pause in public. Jensen Huang was reported as unconvinced.
So what
Treat this as a roadmap risk rather than a philosophical debate. If the labs coordinate any slowdown, capability improvements you have assumed in next year's plan may arrive late, while the cheaper models you already deploy keep getting better. Build your 2027 product commitments on the model tier you can buy today, and keep frontier capability in the upside case rather than the base case. That is prudent whether the pause happens or not.
Hype check
"Open to slowing" is a statement of willingness reported second-hand from an internal meeting, not a commitment, a timeline, or a threshold. It also lands while OpenAI is rationing $200 Pro seats for capacity reasons, which makes restraint and scarcity harder to tell apart.
Confidence: Watch this — the reporting is credible but the statement is directional, with no announced mechanism.
Go deeper Bloomberg · The Deep View
Business
September 10: Miro sold for $1.36 billion, 90% below its 2021 mark
Bending Spoons agreed to acquire the collaboration platform Miro in an all-cash deal valuing the business at $1.355 billion, against the $17.5 billion price it carried in late 2021. Miro is not a broken company: it reports roughly $600 million in annual recurring revenue, nearly 90% from business and enterprise customers, about 4 million paying users, and more than 750 accounts above $100,000 in ARR. Some Miro shareholders agreed to reinvest $295 million of proceeds into new Bending Spoons shares. The deal is expected to close in the fourth quarter.
So what
A durable $600 million ARR business cleared at roughly 2.3 times revenue, which is the clearest public mark yet on what 2021 private valuations are actually worth. For founders and boards carrying a stale preference stack, the exit math has changed and the acquirer pool now skews toward disciplined consolidators rather than strategics paying for growth. Re-run your waterfall against a 2 to 3 times revenue outcome this quarter, and find out which of your employees' options are still worth something. Better to know now than during diligence.
Confidence: Established fact — Bending Spoons published the definitive agreement and filed it with regulators.
Go deeper Bending Spoons · TechCrunch via TLDR Founders
Perspectives
Business
The SaaS collapse repriced expectations, not revenue
Stripe's economics team tracks weekly pay-in volumes across 72,000 non-AI SaaS businesses, and their index tells a different story from the roughly $2 trillion of software market capitalization that evaporated earlier this year. Year-on-year revenue growth sits above 30%, at or above the highs of summer 2023, and the index level is now 4% above its pre-selloff trend. Growth was strongest among younger companies and in the US market, with healthcare, retail and professional services outperforming. The authors argue the selloff priced a forecast about AI substitution rather than observed activity.
So what
The gap between your operating reality and your valuation multiple is currently a gap in narrative, and narrative is something you can address directly. Put transaction-level evidence in front of your investors: cohort retention, seat expansion, and the share of revenue from customers who also bought AI tooling and stayed. Boards repricing you on sector sentiment should be answered with your own data. Do not restructure a healthy business around a multiple that reflects someone else's forecast.
Confidence: Informed perspective — the underlying payment data is real, but Stripe's merchant base is not a representative sample of enterprise software.
Go deeper Stripe Economics · SaaStr via TLDR Founders
People & Organization
The wait was doing work nobody budgeted for
Every senior editor Jack Cheng noticed something after frontier models got efficient enough to run as daily drivers: his weekly usage limits stopped binding, and his judgment got worse. When he had to wait for the limit to reset, questions had time to develop and weak ideas died on their own. Now he ships a feature request to the model the moment he thinks of it, and does what engineering leader Will Larson calls snacking, choosing easy low-impact work over hard high-impact work. He traces the pattern back to time-sharing, quoting Sherry Turkle in 1984 on how hard it is to walk away from a machine that answers immediately.
So what
As you remove cost and latency from execution, you also remove the friction that used to filter your roadmap, and nothing in your process replaces it. That shows up as a backlog of cheap shipped features nobody asked for and a strategy nobody revisited. Reintroduce the filter deliberately: require a one-paragraph written rationale before any agent-built feature enters a sprint, and hold new ideas overnight. The constraint you lost was free; the one you install costs a day and is worth it.
Confidence: Informed perspective — this is one practitioner's account, argued well and consistent with the historical record he cites.
Go deeper Every · Will Larson on snacking
Resources
Tools
September 10: ChatGPT Work got a data agent that answers questions about your own business
OpenAI shipped a data plugin for ChatGPT Work that connects approved company data sources and answers questions in conversation, then builds interactive dashboards that can be shared. It reads your semantic layer, so it uses your business terms, custom calculations and defined metric relationships rather than inventing its own. OpenAI published example prompts for diagnosing a metric change, and its CTO of applications described how the internal data team uses it. The same day, OpenAI launched ChatGPT for Financial Services, an edition of Work with PitchBook, Crunchbase and LSEG data built in for modeling and pitchbook workflows.
So what
The bottleneck this removes is not query writing, it is the queue in front of your analytics team, and the constraint it exposes is whether your semantic layer is any good. Companies with defined metrics will get useful answers; companies where three teams define "active customer" differently will get three confident wrong ones. Pick one recurring executive question this month, connect only the sources it needs, and check the answer against a hand-built number before you widen access. The audit is the adoption plan.
Confidence: Established fact — both products are launched and documented by OpenAI, and available to Work customers now.
Go deeper OpenAI · Superhuman AI
Go-to-Market
September 9: Google made incrementality testing free and publisher-agnostic
Meridian GeoX reached general availability globally on September 9, moving from the May preview to a supported open-source release. It runs geographic experiments by holding a campaign live in some regions while pausing or boosting it in others, then measuring the difference against what historical data predicted anyway. It supports multi-cell tests against a common control group, works across platforms rather than only Google media, and combines online and offline sales in the same test. Results feed back into Meridian's marketing mix model to recalibrate budget allocation.
So what
Platform-reported ROAS tells you which channel claimed the sale, not whether the sale would have happened anyway, and the cost of finding out has now dropped to engineering time. This matters more as AI answer engines and agent traffic scramble the referral data your attribution model depends on. Run one geo holdout on your largest paid channel this quarter, before renewing that budget line. If the incremental number lands far below the platform's, you have found your cheapest margin improvement of the year.
Confidence: Established fact — the code is public on GitHub and Google documented the general availability release.
Go deeper Google Meridian · Search Engine Watch via TLDR Marketing
