Recognise the architecture: a transit station is just a reverse proxy you point base_url at. The open-source gateways behind them (one-api, new-api) are neutral and legal; the risk lives entirely in how the operator sourced the upstream keys. Rule of thumb — if the price is far below what the tokens cost the reseller upstream, someone is eating that gap through fraud or model substitution, and it may end up being you.

The grey market is a signal about pricing arbitrage and enforcement, not an opportunity. It exists because the gap between subscription, free-tier, and API pricing is large — and it is now drawing criminal enforcement (a Shanghai operator was detained in May 2026) and vendor KYC crackdowns. The durable read: tighter identity verification and regional access controls will keep raising the floor under legitimate pricing.

If anyone on your team is sourcing model access through a discount reseller to hit a budget, treat it as a procurement and compliance issue, not a clever hack. The savings are not real once you price in accuracy risk, data exposure, and account bans. Route spend through the legitimate cheap paths — open weights, serverless endpoints, official cloud resale — and make the policy explicit.

Re-run your routing eval: work you escalated to Opus for quality may now clear on Sonnet 5 at roughly 40% of the input cost. Mind the promo — introductory pricing expires August 31, so bake the post-intro $3/$15 into your unit-cost math, not the launch rate. A better cheap tier only saves money if your router actually sends the easy traffic to it.

The competitive move here is mid-tier price compression, not a new frontier. The categories that benefit are anything whose unit economics are gated by inference cost — agent products, high-volume classification, RAG at scale. Watch the intro-to-standard price step (Aug 31) as a pattern: promotional inference pricing is becoming a customer-acquisition lever, which makes headline prices a weaker read on true cost.

A cheaper capable model is a re-baselining prompt, not a set-and-forget. If you priced a feature against Opus tokens six months ago, the model you should run — and the price you should charge — may both have moved. Put a quarterly model-and-price review on the calendar rather than discovering the gap on an invoice.

This is the empirical case for human-in-the-loop and hard bounds on agent autonomy. Models that ace SWE-bench and terminal benchmarks still fail interactive reasoning in novel environments — so design agents to fail safe, checkpoint with a human on unfamiliar tasks, and never infer general reasoning from a narrow benchmark win.

The durable read is that capability is jagged, not monotonic: a model can top coding leaderboards and score near zero on interactive reasoning. Theses that price in near-term general autonomy are exposed; the categories that hold up are narrow, well-scoped agent deployments with verification, not open-ended autonomy plays.

Before greenlighting an autonomous agent for a novel workflow, ask which benchmark the confidence rests on — static-knowledge scores overstate autonomous reliability. The cheap test: can a human verify the agent's output, and what is the cost when it is confidently wrong.

eve standardizes the parts most teams hand-roll — durable execution, sandboxing, HITL approvals, tracing, evals — behind one deploy command, with agents as version-controllable file directories. Worth evaluating against your current stack, but weigh runtime lock-in: the convenience layer is also the thing you become dependent on. Keep agent logic portable.

A major infra vendor making a platform bid for the agent runtime — build, deploy, observe, govern — is the category maturing toward consolidation. The "agents as files + one deploy" framing plus an IdP layer (Passport) targets the shadow-AI control gap enterprises are now feeling. The thesis is the runtime-and-governance layer; usage stats here are vendor self-reported.

Passport is the operator-relevant half: putting employee-built agents behind your existing identity provider is a direct answer to ungoverned "shadow AI" proliferating inside the org. If teams are already shipping agents you have not inventoried, an IdP-gated runtime is one way to regain control — assess it alongside your access-governance plan.

For cost-bound workloads, benchmark a current open-weight model on a serverless endpoint before assuming you need a flagship — the quality gap on many narrow tasks is small and the price gap is large. Buy open weights from a reputable host or self-host at real utilisation; do not source them through a discount reseller you cannot vet.

The open-weight frontier is increasingly Chinese-led and moving fast, which pressures closed-model pricing from below. The category exposed is anything selling undifferentiated inference at flagship margins; the beneficiaries are the serving and hosting layer and products that turn cheap open weights into reliable output.

Open weights change your build/buy math but not your due-diligence obligations — license terms, provenance, and where an endpoint runs your data still matter. The cheap-token temptation is real; make sure the savings come from a legitimate host, not a reseller routing your prompts and keys through an unknown intermediary.

If you prototyped agents in Agent Builder, plan the migration to the Agents SDK now, not in November. The broader lesson: prefer the code-first layer you control over the vendor's visual canvas for anything you intend to run in production — the canvas is the part that gets deprecated. Pin SDK versions and keep your eval harness portable.

The fast reversal validates the thesis that durable agent value sits in code-first frameworks and the surrounding tooling (eval, observability, gateways), not in no-code builders that labs ship and retire. The category that benefits is vendor-neutral agent infrastructure; lock-in to a single lab's convenience layer is the risk being repriced.

Audit which of your agent workflows depend on a vendor's managed builder versus a portable SDK. Anything on the managed layer now has a calendar deadline. Treat "which agent platform are we married to, and what is its deprecation history" as a procurement question this quarter.

More parallel subagents means more tokens — Anthropic's own data puts multi-agent systems at ~15x the tokens of a chat turn. Before you adopt dynamic workflows for a real migration, wire per-run token/dollar budgets, no-progress detection, and KV-cache-friendly (append-only, byte-stable prefix) context. The capability is real; so is the bill if the loop is ungoverned.

Frontier labs competing on sustained agentic compute rather than single-shot benchmark scores reframes the cost curve: the unit that scales is parallel subagent spend. The categories that benefit are agent-cost governance, eval-in-production, and the gateway layer that caps and routes that spend — the picks-and-shovels of the agent era.

A model that can orchestrate hundreds of subagents can deliver a quarter's migration in a session — and can also burn a quarter's token budget in one. Pair any pilot with a hard spend cap at the gateway, because vendor budgets are notifications, not enforcement.

Instrument what you ship. Per-feature unit economics — calls per request, tokens per call, retrieval ops, telemetry overhead — beats per-model dashboards. The three controls that actually move the bill: prompt caching, model routing to smaller models when sufficient, and structured outputs.

AI cost-optimization is becoming a real data-engineering category — FinOps for AI, prompt-caching middleware, model-routing platforms. These are picks-and-shovels for the second half of the deployment cycle, when the bill arrives and procurement starts asking.

If your AI feature P&L looks flat year-over-year while volume grew 5x, you do not have a cost problem — you have a measurement problem. The bill is compounding silently. Get unit economics on the dashboard this quarter.

Before launch, run the six-question audit: who owns this when it breaks; what is the rollback path; how is quality measured continuously; what are we claiming externally; what is the worst-case trust scenario; is leadership ready for the press cycle. Most teams have not answered three of those.

The rollback list is the buyer-feedback signal. Vendor themes that match: pricing-model design for agentic workloads, eval-in-production, claim-calibration tooling. These categories have a story buyers now believe.

Calibrate external claims to internal capability. The Klarna pattern — externally claim 100%, internally have 70% — sets up the press cycle. Underpromise during launch; let outcomes do the marketing.

Agent-platform optionality is becoming a real architecture decision: Anthropic, OpenAI, and Google are each pushing their own orchestration layer. Keep your agent logic in a portable layer and treat each lab's platform as a swappable backend, so a frontier shift does not become a rewrite. Verify capability claims against your own evals before trusting the keynote framing.

Three frontier labs converging on multi-agent orchestration is a category-formation signal: the value accrues to the orchestration, eval, and cost-governance layers that sit above any one model. The risk is over-indexing on keynote claims — published adoption numbers from this cycle were thin enough that one circulated protocol-adoption figure could not be primary-sourced.

If your roadmap assumes a single agent platform, the I/O signal is that the ground is still moving — budget for portability. The orchestration race is good for buyers (more capability, falling prices) but raises the premium on not being locked into one lab's agent stack prematurely.