Executive Summary
Today's sources agree that model capability is no longer what limits enterprise AI. They disagree on where that leaves the model layer. Nate Jones argues that leaderboard rank has stopped predicting commercial value and that product fit decides who wins. Matthew Berman shows the operational version of that claim: reliable agents built from narrow scope, explicit permission boundaries, and approval gates, with no particular frontier model required. Simon Willison's teardown of Claude Haiku 5.5 is the counterweight. Model choice still matters, but the price sheet tells you little. Tokenizers, context-length pricing tiers, and reasoning defaults now decide unit cost. The practical stance: design agent systems to be vendor-agnostic and governed by architecture, then pick models per workload using measured token economics.
What Changed
- Haiku 5.5 matches GPT-6 Luna on list price ($0.10/$0.50 per million tokens) up to 100K tokens. Above that, Haiku jumps 5x to $0.50/$2.50. Luna holds its base rate until 272K and then rises only to $0.20/$0.75.
- The new Haiku tokenizer uses about 1.25x more tokens than Haiku 4.5 for the same prompt. Effective short-context cost works out to roughly $0.125/$0.625. List-price parity holds on paper only, unless Luna's tokenizer is equally inefficient (not established).
- Haiku 5.5 can no longer run without reasoning. It defaults to medium effort. Willison's test ranged from 7 seconds at about $0.001 (low effort) to over 5 minutes at $0.034 (max effort). The cheap, fast, non-reasoning path is gone.
- Max and Team subscribers now get monthly API credits matching their subscription cost (Team up to $500 pooled). The credits do not roll over.
- Google's Gemini Argon tops a leaderboard but has no clear access path or place in Google's products. Jones reads this as a business weakness, not a technical one.
Cross-Expert Synthesis
Berman and Jones arrive at the same architecture from different directions. Berman's twelve agents each get one job, one data source, and a standing "never do X without my okay" clause. They coordinate through shared state (one to-do list, one calendar) with no orchestrator. Jones runs a slow, connector-heavy tool for long-horizon aggregation alongside a fast tool for execution. In both cases reliability comes from decomposition and scoping, not from model intelligence. That supports Jones's thesis better than his Argon example does.
The tension is with Willison. If capability has stopped being the constraint, buyers will treat models as interchangeable and route on cost. Willison shows that cost has become opaque in ways that punish exactly that buyer. A team that moved high-volume classification to Haiku for speed now pays for reasoning it cannot turn off, on 25% more tokens. "The model doesn't matter" holds for capability ceilings. It is wrong for unit economics. Vendors now compete on billing mechanics more than on benchmarks, and procurement is not set up to see it.
Enterprise Implications
Berman's governance pattern costs little and can be inspected. System prompts set read and write boundaries and require draft-then-approve for sends, spending, bookings, and deletions. Trust is what blocks most internal agent rollouts, and this pattern lets a security or risk team review a deployment by reading its prompt. The reported ROI is replicable: a subscription audit agent recovered over $1,000 a month across three accounts, and a rate-comparison agent saved about $1,000 a year. Those map directly onto spend reviews in accounts payable and procurement.
The cost of shared state is the underexamined side. Agents that coordinate through one to-do list or calendar have no central place to resolve conflicts or keep an audit trail. That works for one person. Across a department it turns into a data-integrity problem.
Pricing tiers by context length make RAG and document-heavy pipelines especially sensitive to vendor choice. At 100K to 272K tokens, Haiku is roughly 5x Luna on list price and more after the tokenizer adjustment. Any workload with long retrieved context should be benchmarked on real traffic before committing.
What To Do About It
- Re-baseline model costs on token counts from production prompts, not list prices. Run a representative sample through each candidate's tokenizer, segmented by context length.
- Set reasoning effort explicitly on every Haiku 5.5 call. Audit any pipeline that picked Haiku for latency, since the medium default changes both its p95 latency and its cost.
- Make the approval gate a deployment requirement for agents. Any irreversible action (send, spend, delete, book) needs a "draft and wait for approval" clause written into the system prompt, checked at review time, not assumed.
- Claim the Team and Max API credits now and disable auto-reload. The credits don't roll over, and turning off auto-reload caps spend.
- Drop leaderboard rank from vendor scorecards. Score instead on access path, how well the vendor's customer focus fits yours, and measured workload economics.
Sharp customer question: "When your AI vendor changed its tokenizer last quarter, who in your organization noticed?" Most enterprises will have no answer, and that gap is a consulting engagement.
Risks and Blind Spots
Berman's ROI figures come from personal accounts. They show the pattern works but don't establish enterprise savings, and renegotiation drafts sent at scale could damage vendor relationships. Approval gates that live only in prompts are a policy, not a control. A prompt injection through an ingested email could argue the agent past its "never without approval" clause. Production use needs the gate enforced at the tool and permission layer.
Jones's prediction that Fable 5.5 will beat Argon on enterprise value rests on how Anthropic is organized, not on evidence, and his own multi-tool workflow is one person's setup. Price comparisons today assume Luna's tokenizer efficiency, which none of the sources measure. Vendors can change tokenizers, reasoning defaults, and pricing tiers without notice, so any cost model built this quarter has a short shelf life. Re-measure every quarter.