Harness Pricing Models Compared for Engineering Teams

Harness pricing models determine how engineering teams pay for agentic work, but cost per completed task is the soundest basis for comparing them. For a bring-your-own API key cloud coding agent runner, the comparison must distinguish model charges, harness charges, and sandbox runtime. The same distinction applies when evaluating cloud execution of Claude Code or Codex without managing infrastructure.

Why harness pricing has become a first-order budget problem for engineering teams

Uber capped AI coding spend for each engineer after monthly costs differed widely from one engineer to another and a single two-hour CTO demonstration spent a large sum on tokens. The Microsoft case shows a different risk, where flat seat pricing hid how much each session used until the total costs came into view. Its root cause was structural.

Line completion has given way to agents that read an issue, edit files, run tests, and open a PR. Those loops consume dramatically more tokens than a chat window, while the amount of work inside each session varies. A billing structure that was barely visible during a pilot becomes an operational constraint when engineers run agents throughout the day.

The three billing structures engineering teams are actually choosing between

Flat seat licenses charge the same amount each month for each developer, however much they use. The purchased unit is access for a person. This makes the subscription straightforward to budget. It does not, however, expose the cost of individual sessions, so the license total and the economics of the work remain separate views.

With tiered subscription caps, that monthly seat fee comes with a usage allowance. Once the allowance is exhausted, further usage is either blocked or billed separately. These plans commonly express the allowance as a multiplier on the entry tier, with Claude Code Max and Cursor Ultra offering substantially higher limits. The multiplier describes entitlement. Teams therefore need to compare how much usage is included, even when products have similar entry and top-tier plans.

With pure token consumption, there is no harness subscription to pay for. OpenCode is free and open source, and teams pay for underlying model tokens at the provider’s published rates without a harness markup. Linear’s coding sessions illustrate a hybrid as of October 2026, drawing from a prepaid workspace-level pool of AI credits. Model tokens are charged at published provider rates without markup, while sandbox runtime is billed per timed block. That distinction separates the cost of inference from the cost of the environment running the agent.

How agentic workloads break the assumptions behind each billing structure

A single agentic task can trigger hundreds of recursive model calls as the harness spawns subagents, runs tests, and retries failures. Context grows as the loop continues. The third retry can cost more than the first because the request includes the history of earlier failures, so counting retries alone understates consumption. Uber’s wide variation in spending is the operational signal that seat count cannot capture this workload distribution.

Under flat seat pricing, the price developers see hides those differences until the total usage comes into view, as the Microsoft case shows. The fixed license does not describe the underlying workload. Under capped subscriptions, the same amplification appears as depleted allowances instead. A developer running background agents can exhaust a monthly allowance in days, leaving the team to accept interrupted work or move seats into a higher tier.

Pure token billing exposes consumption but does not make it predictable. Without a cap, a runaway loop or a session that uses far more tokens than usual can run up charges that a forecast based only on headcount would miss. Transparency is not the same as control. The risk shifts from whether the allowance will last to how much the metered usage will cost.

All three structures also share a context re-loading cost when the agent does not retain memory across tasks. Each new session then requires it to read the codebase context again, while repeatedly sending full history costs more than using pruned context. The agent spends less on repeated input and avoids being overwhelmed by fresh context on every turn.

What cost per completed task reveals that cost per seat and cost per token both hide

Cost per completed task captures the resources consumed in reaching an engineering outcome, rather than the price of access or an isolated model call. The Composio benchmark used the same underlying model across harnesses, isolating harness efficiency from differences in model capability. That methodological choice is essential to the comparison. Two harnesses can run the identical model on the identical task yet consume significantly different totals because their scaffolding, loops, and retry behavior differ.

Control the Harness, Control the Cost Routing and Governing AI Coding Agents in the Enterprise found that cache reads dominate the bill in long, tool-heavy sessions. At Anthropic’s September 2026 list prices, Fable 5.1 re-reads cached context at half the price of Opus 5.5, despite being the higher-priced model. In a long coding session that repeatedly reads cached conversation history between tool calls, that lower cache-read price can outweigh its higher price elsewhere.

The routing study found this crossover in a majority of real sessions above a cost threshold. The mix of fresh input, cached input, and other session work determines the result. Seat and token prices remain useful inputs. Neither establishes the total cost of completing the task without evidence about the harness loop and the session’s consumption pattern.

How each major harness structures its costs at team scale, and where the spikes happen

Claude Code’s pricing as of September 28, 2026 includes Pro, a mid-tier Max option, and a top-tier Max option, with Team Standard and Team Premium seats scaling by usage allowance. Team Premium runs $100–$125/seat per month. Default routing uses Anthropic’s models, while enterprises using the API can redirect requests through a custom endpoint to other vendors’ models. Teams considering high-limit seats should establish whether those limits cover their actual agentic sessions.

OpenAI Codex has a published top Pro seat price significantly above the market norm, positioning it for high-volume agentic use cases that would overflow other harnesses’ caps. That price needs to be evaluated against the work included in the comparison. Repository-scoped coding is not the whole workload. Codex’s “Work mode” covers broader deliverable tasks that consume tokens differently, so teams should benchmark both modes against their expected mix before treating the seat price as a forecast.

Cursor’s X Subscription plans, launched October 12, 2026, use a weekly pool for Cursor models including Grok 4.7, 4.6, 4.5 and Composer 2.5. Legacy Pro, Pro+, and Ultra subscribers keep their monthly pools. Background runs can drain a weekly pool mid-sprint, so a team with concentrated usage around release deadlines faces a different constraint from one whose work is spread evenly. As of October 18, 2026, the plans include Free, Plus for Grok models only, Premium, Super $100, Ultra, and Teams.

OpenCode places model selection and session volume directly on the bill, with no harness markup and support for any models.dev provider. Its flexibility makes provider prices visible without a bundled seat allowance. Without a cap, a runaway agent loop has no built-in spending bound, so teams need their own controls at the API or infrastructure layer. Its cost-per-task comparison therefore depends on which model the team chooses and how the harness runs the team’s work.

Claude Code and Codex require attention to seat entitlement, Cursor adds reset cadence, and OpenCode makes metered model consumption explicit. Hosted execution can add another charge. Linear’s separate sandbox runtime billing illustrates why a cloud agent comparison must account for execution as well as tokens, even when inference itself has no markup.

Model routing as a practical lever for controlling harness spend

The largest controllable variable in harness spending is which model handles a task and when that assignment occurs. The routing study emulated a large enterprise using behavior drawn from real session datasets and recovered a significant fraction of annual model spending by choosing the cheapest model capable of each task. The savings depended on preserving work in progress. Even a one-line change in a long coding session could cost meaningfully less when routed to the appropriate model at the appropriate moment.

The practical rule is to move work at session boundaries. A mid-task switch discards the cached conversation and requires a full context rebuild, which must be repaid through enough subsequent requests to justify the change. Corrections demand a different safeguard. Teams should not downgrade a correction, because assigning it to a model that is too weak can trigger further corrections and compound the bill.

Pause behavior also belongs in the routing calculation because the API cache has a short expiry window. A significant share of user turns follow pauses long enough for the full conversation to be billed again. Engineers who leave longer gaps within an active conversation can incur more cache-rebuild costs than those whose requests arrive close together, so routing cannot be evaluated independently of working patterns.

Custom base URL endpoints give teams using Claude Code and OpenCode access to other vendors’ strong models at different price points. Teams can make that choice for whole task categories as well as for single prompts. A multi-harness platform supports that broader choice. Rather than bind every task to one vendor’s loop, a team can route categories to the harness and model combination with the best measured cost per completed task.

What to measure before committing to a harness pricing structure

The purchasing sequence should begin with the team’s workload, then map that evidence to a pricing structure. Session length distribution distinguishes short interactive exchanges from long background tasks, while task complexity establishes what the agents must complete. Those are different demands on the same allowance. Capped tiers work better for shorter interactive use, while token billing needs spending controls however long the sessions are.

The Composio benchmark shows why a representative task mix is necessary, since simple audits and repository sweeps have different efficiency patterns from complex multi-system workflows. Pause behavior belongs in that measurement too. Long breaks inside sessions can force cache rebuilds, adding costs that a seat tier does not reveal and changing the economics of model routing.

Cost per completed task should be recorded with attribution at both the harness and model level, because routing decisions cannot be evaluated without knowing which run incurred which cost. Rework hours per merged PR offer a separate way to measure results and can show what defects cost beyond the agent bill. Completion alone is not sufficient evidence of value.