A per-session cost profiler for coding agents: it attributes token spend to the work being done (which files, which tool calls, which re-reads), identifies the waste ceiling that tool-output bloat and repeated context create, and applies measurable interventions — output summarisation, retrieval instead of full-file reads, model routing per task type — then reports the verified saving. The bill is already visible; the accounting is what is missing.

Market Panorama

Competitive Intelligence

Users & Demand

Business Model

Market Sizing

Tech Solution & Build Plan

Go-to-Market Strategy

Risk Assessment

Action Recommendations

### Market definition and scope

The cost optimiser sits on the demand side of inference economics. It captures what a coding agent spends per turn, tool call and file read; attributes that spend to files, tasks and projects; identifies the recurring waste patterns — repeated file reads, oversized tool output, retry loops, context re-sending — and applies interventions through harness hooks rather than by forking a harness. Each intervention is measured against a baseline so the saving is a verified number rather than a claim.

The boundary matters because the adjacent space is crowded in a specific direction. It is not an inference provider, not a router, and not a model. It is not application-level LLM tracing either: LangFuse and LangSmith instrument applications whose developers control the call sites, whereas a coding agent is a third-party binary that a founder runs interactively and cannot instrument from inside.

Its unit of value is a screenshot: "you spent $312 this month and $118 of it was re-reading files you had already read."

That is the entire product thesis, and October's data supports both halves of it. The supply side is validated and crowded. deepseek-ai/DeepInfer took 25,086 stars and 2,355 points with 892 comments (GitHub Trending; Hacker News, October 2026) for making DeepSeek inference faster. Zagitta/gpu-unlock reached 14,928 stars selling up to 2× tensor throughput on consumer cards. microsoft/TokenEconomizer reached 4,557 stars for black-box cost control over coding agents. Show HN's Llmspeed reached 331 points and 86 comments helping people feel how fast a local model runs before buying hardware.

### Key trend signals with data citations

The supply-side concentration is itself the signal. Four independent projects in one month, all attacking the cost of inference from the production end, means cost is acknowledged as the binding constraint. None of them answers the question a founder asks when the bill arrives: where did it go?

The pricing substrate arrived in the same window. DeepSeek published per-request search and fetch pricing at $0.75 per 1,000 successful requests (607 points / 282 comments). Granular unit pricing is what makes optimisation actionable at small scale, because it converts a vague "tokens are expensive" into "this pattern costs $18 a month".

TokenEconomizer's framing is the most useful single citation. A Microsoft project describing coding-agent cost as a black box is an admission from a well-resourced team that the accounting does not exist — and it is a research-grade controller rather than a product a founder installs and trusts.

### Why now

Two conditions make this possible and one makes it urgent. Harnesses now expose hooks and proxy configuration, which means non-invasive capture works without forking. Per-request pricing made small waste measurable in money. And the volume of agent usage crossed the threshold where a founder notices the bill and cannot explain it.

The reason the demand side stays empty is structural: a harness vendor will ship a cost display, but a harness vendor will not recommend routing work away from its own model or cutting the context that makes its product feel capable. The intervention layer is the part vendors cannot credibly sell.

### Market maturity

Growing, with an unusual split. The supply side is genuinely competitive with well-funded entrants; the demand side has no founder-installable product. That asymmetry is what makes the opportunity available at all, and it is also why the scoring is cautious — a growing market draws entrants.

### What this means for a solo founder

Ship the free report first and let the number sell the product. The supply side is crowded and the accounting is empty; the entire business is a verifiable sentence about where the money went.

deepseek-ai/DeepInfer took 25,086 stars and 2,355 points / 892 comments on Hacker News — the month's most-discussed technical launch.

Inference optimisation goes mainstream

Zagitta/gpu-unlock reached 14,928 stars advertising up to 2× FP16/BF16/TF32 throughput on 20/30/40-series cards.

Consumer hardware performance as a product

microsoft/TokenEconomizer reached 4,557 stars for black-box cost control over coding agents.

Coding-agent cost named as a black box

DeepSeek priced search and fetch at $0.75 per 1,000 successful requests (607 points / 282 comments), making unit-price optimisation actionable.

Coding-Agent Token Cost Optimiser