A light, end-of-month day with the action in tooling economics, not models.
Two AI-coding pricing dominoes land at once — Cursor's Teams renewals re-price tomorrow (Jul 1) and GitHub Copilot's first full metered billing cycle closes today — making "what does our agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ dev tooling actually cost" the question worth an hour this week. Separately, Anthropic's science push gets its own showcase (AI for Science livestream today), and the Google talent story gains a concrete structural twist: a reorg of its coding "strike team" into midtraining.
TLDR
- Cursor Teams pricing re-prices on renewal starting tomorrow (Jul 1): split Composer/Auto vs third-party-API usage pools + a new Premium seat (announced Jun 1, renewals Jul 1) Pricing
- GitHub Copilot's first full usage-based billing cycle closes today; power users report 10–50× bill jumps on agentic sessions (cycle close Jun 30) Pricing Trending
- Anthropic "The Briefing: AI for Science" livestream today (10am PT), spotlighting Claude in life sciences (Jun 30) Trending
- Google reorgs its AI coding "strike team" to add a midtraining phase as a 6th researcher (Denny Zhou) surfaces at Meta (reported Jun 25–28) Hiring
Action Items
Things your engineering team should look at or do this week.
- Audit Cursor Teams seats before your renewal (re-prices Jul 1). Standard seats now split into two pools — Composer/Auto (first-party models) and Third-Party API (Claude/GPT/Gemini) — and there's a new Premium seat (5× included usage at 3× cost: $96/seat/mo annual, $120 monthly). Map your heaviest agent users to Premium and everyone else to Standard before the new structure hits at renewal. (Cursor)
- Reconcile your first GitHub Copilot metered bill (cycle closes today) and set spend controls. Code completions/Next Edit stay free, but agentic sessions, premium-model calls, and PR review now burn GitHub AI Credits ($0.01 each at per-model API rates), and the old cheaper-model fallback is gone. Pull the org usage report, identify who's running agent mode all day, and decide on caps before next cycle. (GitHub)
- For agent builders: put deterministic tools under the model, not bigger models on top. Anthropic's biology work found agents hit ~100% on a retrieval benchmark only after adding a deterministic data-access layer — a portable lesson for any RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗/agent pipeline where reliability matters more than the newest model. (Anthropic)
APIs & Pricing
- Cursor overhauls Teams pricing; new structure hits on renewal from Jul 1 (announced Jun 1, renewals Jul 1) Pricing — Every Teams seat now carries two separate included-usage pools (Composer/Auto for first-party models; Third-Party API for Claude/GPT/Gemini), with a new Premium seat for heavy agent users (5× usage at 3× cost — $96/seat/mo annual, $120 monthly; Standard is $32/$40). Changes apply immediately for new customers and at renewal for cycles starting Jul 1, 2026. Cursor says ~90% of teams pay less; the catch is that all-day agent users need the Premium tier to stay predictable. (Cursor)
- GitHub Copilot's first full metered cycle closes today amid bill-shock reports (cycle close Jun 30) Pricing Trending — Since usage-based billing went live Jun 1, the first complete monthly cycle ends today, and developers are posting screenshots of projected costs jumping 10–50× (e.g., $29→$750, $50→$3,000) on heavy agentic workflows. Mechanics: 1 GitHub AI Credit = $0.01, metered at per-model API rates; completions/Next Edit remain free; a single agent session can run $30–$40; annual plans are being retired. The broader signal for tooling budgets: the flat-rate "unlimited AI" era is ending across coding tools. (GitHub)
Industry & Business
- RAISE US launches with $500M+ to retrain US workers for the AI economy (Jun 25) Policy — A new nonpartisan nonprofit from ex-Commerce Secretary Gina Raimondo and ex-Indiana Gov. Eric Holcomb, targeting $1B in commitments and already past $500M, with Amazon, Anthropic, Microsoft, and the OpenAI Foundation as anchor partners. It funds state-level retraining/apprenticeship pilots tied to documented employer demand. Notable for the AI vendors themselves co-funding workforce transition — and an honest framing that prior retraining programs largely failed to move workers into less-automatable jobs. (Rockefeller Foundation | Axios)
- Google reorgs its AI coding "strike team" into midtraining; 6th DeepMind researcher surfaces at Meta (reported Jun 25–28) Hiring — Structural update to the Google brain-drain story. Per The Information, DeepMind expanded its April-formed coding strike team (led by Sebastian Borgeaud, overseen by Brin and Kavukcuoglu) to add a dedicated midtraining phase aimed at lifting Gemini's coding ability at the model level rather than via tooling/post-training. It coincides with a sixth named departure becoming public — reasoning-research founder Denny Zhou, already ~4 months into Meta's Superintelligence Lab — extending a run of exits across reasoning, architecture, science, coding, and pretraining. (Neowin | TechTimes)
Models & Tools
- Anthropic "The Briefing: AI for Science" virtual event today (Jun 30) Trending — A livestream (10am PT) with Anthropic leadership, life-sciences executives, and research institutions on Claude's use across science — the first big showcase since Nobel laureate John Jumper (AlphaFold) joined from DeepMind (announced Jun 19). It builds on Anthropic's June biology research, which introduced the VirBench retrieval benchmark and the deterministic gget virus tool (built with NCBI) that pushed model accuracy from as low as ~17% to ~93–100% — a concrete data point that reliable agent behavior comes from boring deterministic tooling, not just stronger models. (Anthropic | CNBC)