Executive Summary
The model is no longer the scarce input in enterprise AI. Today's sources point to three bottlenecks that have replaced it: choosing workflows, building evaluations, and governing what agents create. Frontier-adjacent capability now costs two to three orders of magnitude less per task. Anthropic's S-1 shows even its largest enterprise customers spend under 1% of total expenses on it. A growing share of vendor and architecture decisions is being made by models, benchmark vendors and agents rather than buyers. Enterprises that own their eval harness, routing layer and agent-governance controls will capture the cost collapse. Those that don't will have their stack chosen for them.
What Changed
- Price per task collapsed while volume grew. Opus 5.5 cut cost per task about 40%. DeepSeek V4.1 Flash (open weights) is quoted at about $0.26 per completed task, against $763 for Claude Fable 5.1 Max. Ramp shows token volume up 52% as per-token prices fall, and about 90% of September agent tokens were cached. Cheaper intelligence is increasing total consumption, not capping it.
- Anthropic disclosed shallow penetration and strong margins. It reported H1 revenue of $16.3B, implied gross margin near 85% (excluding R&D) and $974M of Q2 operating profit. Its average public Fortune 500 customer still spends under $1 per $1,000 of expenses.
- Taste became automatable. In Willison's informal tests, a bare "make it a LOT more fun" prompt now works on Opus 5.5. Design judgment, which most organizations assumed needed a specialist, is moving into the model.
- Agents became the buyer. Supabase launches about 5M databases a month, mostly provisioned by coding agents. It says frontier models recommend it first by default.
Cross-Expert Synthesis
Open vs. closed is two separate races. Lambert's read of AAv5 says the gap is widening: eight closed models beat the best open model, Kimi K3.1 Pro. Berman says DeepSeek V4.1 Flash "crushed everyone." Both are right because they measure different things. AAv5 measures the capability ceiling. Berman measures quality per dollar, and DeepSeek's roughly 42 on Humanity's Last Exam and Terminal-Bench puts it level with Kimi K3, not at the frontier. Lambert's view of open models as commodity inference plus ownership of intelligence on internal data, strongest in coding, matches what Berman demonstrated. Lambert also predicts open labs will gain substantially within 2–3 months.
Rankings are unstable and conflicted. Over 80% of Anthropic's revenue comes from models less than six months old. Willison expects OpenAI to match Opus 5.5 on fun by year-end. Lambert notes that index vendors are paid by the labs they rate, and that private benchmarks remove community scrutiny. Any model choice has a shelf life of a quarter or less, and public scores are a weak basis for making it.
Value has moved upstream of implementation. Jones says choosing the workflow, not access to a model, limits Anthropic's expansion. Willison says the scarce skills are now concept and playtesting. Supabase says enterprise adoption comes through greenfield internal tools built by non-engineers, not by migrating production OLTP systems. Execution is cheap. The hard parts are knowing what to build and verifying that it works.
Distribution is becoming model-mediated. When the model picks the backend and a paid index shapes the shortlist, the buyer's formal selection process covers less of the stack than it appears to.
Enterprise Implications
The cost spread forces a routing architecture. Paying closed-frontier prices for high-volume, cache-heavy agent loops will be hard to defend once a Flash-class model handles the same work at a fraction of the cost. Free or near-free cache hits on Lithos and Base10 make long-context agent loops nearly free. Open weights also expose full reasoning and tool-call traces, which strengthens auditability for private and sovereign deployments. Berman ran a 50-minute autonomous coding task entirely on a DGX Station, which makes on-prem near-frontier coding agents plausible.
Agent-provisioned infrastructure is the governance gap nobody has budgeted for. Non-engineers' agents will create databases, apps and compute faster than IT can inventory them. Supabase's Control Tower and its git-native declarative schemas show the controls the market expects: state that lives in git, visibility across the whole fleet, and row-level security by default.
Deployment economics compound. Jones's case for a $100K account becoming $1M rests on reusing the same data connections, systems and validation harnesses for each new workflow. A first deployment that leaves no reusable infrastructure behind is a sunk cost.
What To Do About It
- Treat the eval harness as a core asset. Build task-specific evals on real workloads, rerun them on every major model release, and read several indices with known biases instead of one composite score.
- Build a two-tier routing layer. Reserve closed frontier models for judgment, design and hard reasoning. Send coding, extraction and cached agent loops to open models. Test self-hosting against Lithos and Base10 using your own cache-hit ratios.
- Set the default stack for agents yourself. Put approved backends, schemas and security policies into templates and agent instructions so the model doesn't choose them for you. Inventory agent-created resources now.
- Design the second workflow at the first kickoff. Scope every deployment so the next one is cheaper to add.
- Track spend against value. As reference points, Supabase spends about $670 per head per month, and its host suggests token spend of 0.1–1.0x salary per employee. Measure success by end-to-end processes automated, not tokens consumed.
Customer question worth asking: "Which of your infrastructure vendors were chosen by a coding agent rather than by your architecture review?"
Risks and Blind Spots
- Cost-per-task figures may not compare like with like. The $0.26 vs. $763 comparison likely involves different task difficulty and reasoning effort. Berman's video is sponsored by CoreWeave and the figures are secondhand.
- Most claims come from interested parties. Supabase's numbers are self-reported, including its claim that all tested Postgres operators can lose data, and the interview host is an investor and sponsor. The S-1 is the issuer describing itself.
- Today's prices may not last. Compute scarcity props up open-provider margins (Lambert) and limits Supabase's growth, so current prices could move either way.
- Anthropic's growth may depend on services. The S-1 does not separate software-driven expansion from growth carried by forward-deployed engineers. If the FDEs are a permanent requirement, the expansion economics deteriorate.
- Monitoring that waits for visible symptoms will miss silent failures. Harper's point is that COVID beat containment because it spread before symptoms appeared. Applying that to AI is my analogy, not his claim: silent data leakage, drift and compounding agent errors spread before anyone notices them. AI SRE remains unsolved, by Supabase's own admission.
- Shipping fully AI-built products externally carries reputational risk. Willison reports that audiences already reject AI-generated work, and he would not enter his games in traditional jams. Agents converging on one way of building software (mode collapse) adds a monoculture risk.