AI Model Pricing Splits | AI Field Notes #63

AnalysisThe most expensive model Anthropic sells just left the flat-rate plans. Starting July 8, Claude Fable 5 stops drawing from a subscriber's weekly allowance and bills through separate usage credits: $10 per million input tokens and $50 per million output, double the rate of Opus 4.8. Anthropic pushed the full cutover to July 12 after complaints and called the change temporary, blaming a capacity crunch. For a developer running Fable 5 in an agent loop, a heavy day of roughly thirty runs lands near $37, or $1,100 a month. The subscription that felt unlimited now has a meter on it.

AnalysisSingle-shot coding tests stopped separating the good models from the great ones, so Meta released a harder one. SWE-Together, open-sourced on July 6, runs an agent through 109 tasks that each take several turns of writing, running, and fixing code, then measures how often it lands the fix and how much human correction it needed along the way. Claude Opus 4.8 topped it at 63% solved on the first pass with the least steering. The steering count matters more than the raw score. A model that needs fewer nudges is one a developer can actually leave alone.

AnalysisA year ago Chinese open-weight models were a rounding error on OpenRouter, the marketplace where developers route each call to whichever model fits. Now they carry 45% of its traffic, up from under 2%. Xiaomi's MiMo line alone takes 21.1%, ahead of OpenAI's 7.5%, according to second-quarter data. The pull is price and openness: these models cost little to run and are free to self-host, so a developer watching the bill swaps them in without a second thought. The frontier labs still win the benchmarks. They are quietly losing the default.

AnalysisA ransomware attack ran start to finish with no human at the keyboard, and the security firm Sysdig has the logs. In an analysis published this week, Sysdig documented an AI agent it calls JADEPUFFER that broke in through CVE-2025-3248, a known flaw in Langflow (a popular tool for building AI workflows), then fired more than 600 payloads, hit an error, and rewrote its own approach in 31 seconds before continuing. Automated attacks are old news. An intruder that debugs itself mid-break-in is the part that should keep a security lead awake.

AnalysisThe scarce thing in AI stopped being chips. Now it is a place to plug them in. Anthropic underlined that on July 7 with a long-term lease reported at $19 billion for data-center capacity from TeraWulf, a former bitcoin miner that pivoted its cheap zero-carbon power sites to AI compute. TeraWulf's chief executive called the deal the tip of the iceberg for power demand. A number that large, locked in years ahead, tells you Anthropic expects its models to keep getting hungrier, and that securing electricity now beats scrambling for it later.

AnalysisWiring an agent used to mean gluing together a model client, a tool loop, and your own approval checks. Vercel folded all of that into the AI SDK, its open-source TypeScript toolkit, with the version 6 line updated again on July 6. The release adds first-class agents, a step where the code pauses for human sign-off before running a tool, a debugging panel, and native support for MCP (the Model Context Protocol, the standard for connecting agents to outside tools and data). None of it is novel on its own. Putting it one npm install away from millions of JavaScript developers is the move.

AnalysisWhile the giant chip stories run on billions, a smaller bet just went public. Syntiant, which makes low-power chips that run AI directly on devices like earbuds and sensors instead of in the cloud, filed to list on the Nasdaq under the ticker CBRS. Its filing showed $64.5 million in first-quarter revenue, up 76% from a year earlier, against a $20.9 million loss. The pitch is that not every AI task needs a data center. Some of it should run on a battery, in your pocket, with no round trip to a server.

AnalysisRocket and satellite money may be heading toward your pocket. SpaceX showed investors an early prototype of a consumer AI device, according to a Wall Street Journal report, its first move toward hardware aimed at everyday users rather than launch pads. Form factor and timing stayed unconfirmed, so treat this as a signal, not a shipping product. The interesting part is the pattern: another company with deep pockets and its own satellite network wants a direct line to users, cutting out the phone makers and cloud providers that sit in the middle today.

AnalysisAlibaba folded five separate AI teams into a single group under executive Eddie Wu on July 6, a housekeeping move that says more about scale than org charts. China now processes about 140 trillion AI tokens, the chunks of text models read and write, each day nationally, and Alibaba wants one hand steering how its models, cloud, and apps consume that volume. Consolidation like this usually follows a stretch where internal teams competed and duplicated work. Pulling them together is how a company stops paying twice for the same capability once the bills get serious.

AnalysisThe company most people still treat as the challenger is now the one out front on sales. Anthropic reached roughly $47 billion in annualized revenue as of May, ahead of OpenAI's estimated $25 billion to $33 billion, according to reporting Fortune confirmed this week. Claude Code, its coding tool, drives much of the climb. The gap matters because revenue, not hype, is what funds the next round of compute and hiring. OpenAI still has more consumer reach, but on the money that builds frontier models, Anthropic quietly moved into the lead.