LumeAPIShort path: Cheap LLM API · Cheap OpenAI API · AI API pricing · Usage logs
Last verified: July 26, 2026
A single coding-agent pull request can calculate from about $0.009 for a small mini-model edit to $1.75 for a large premium-model run under the transparent scenarios below. Real cost can be higher because repository reads, tool schemas, test output, retries and failed attempts repeatedly consume tokens. The useful business metric is total model spend divided by accepted pull requests—not the price of one chat completion.
Quick answer
| PR size | Token assumption | Low-cost route example | Strong-route example |
|---|---|---|---|
| Small | 20k input / 3k output | GPT-5.4 mini: $0.0086 | GPT-5.6 Terra: $0.0285 |
| Medium | 100k / 15k | GPT-5.6 Terra: $0.1425 | GPT-5.6 Sol: $0.2850 |
| Large | 400k / 60k | GPT-5.6 Terra: $0.5700 | Claude Opus 4.8: $1.7500 |
These are calculated LumeAPI token charges using prices checked July 26, 2026. They are not observed customer invoices.
In short
A single coding-agent pull request can calculate from about $0.009 for a small mini-model edit to $1.75 for a large premium-model run under the transparent scenarios below.
The pull-request cost formula
For one model route:
PR cost = input_tokens / 1,000,000 × input_rate + output_tokens / 1,000,000 × output_rate
For a real agent with retries or multiple models:
accepted PR cost = Σ(all attempt and model costs) / accepted PRs
The second formula prevents a misleading result. Three $0.10 failed attempts followed by a $0.25 accepted attempt cost $0.55, not $0.25.
Complete model-price matrix
| Route | LumeAPI input/output per 1M | Small 20k/3k | Medium 100k/15k | Large 400k/60k |
|---|---|---|---|---|
gpt-5.4-mini | $0.225 / $1.35 | $0.0086 | $0.0428 | $0.1710 |
gpt-5.6-terra | $0.75 / $4.50 | $0.0285 | $0.1425 | $0.5700 |
gpt-5.6-sol | $1.50 / $9.00 | $0.0570 | $0.2850 | $1.1400 |
claude-sonnet-4-6 | $1.50 / $7.50 | $0.0525 | $0.2625 | $1.0500 |
claude-opus-4-8 | $2.50 / $12.50 | $0.0875 | $0.4375 | $1.7500 |
LumeAPI currently lists up to 70% off selected GPT official rates and 50% off the listed Claude routes in this table. Discounts and model availability can change; verify /models before budgeting.
What drives agent cost per PR?
Repository and diff size
A large repository does not automatically enter the prompt, but repository maps, searches and repeated file reads can make input tokens dominate. Generated files and vendor directories should be excluded from exploration.
Test and terminal output
One failing integration test can produce thousands of lines. If the full output is sent back to the model repeatedly, the debugging loop becomes an input-token multiplier. Truncate logs and extract the first actionable error.
Tool rounds and retries
Each agent round can resend instructions, tool definitions and prior context. Network retries may be cheap if the request was not billed, while semantic retries are new model calls. Measure both.
Model escalation
Starting every typo on the strongest model wastes budget. Starting every difficult refactor on the smallest model can also waste budget through rework. Escalation needs a rule.
Acceptance rate
A cheap generated diff rejected by a reviewer provides no production value. Include accepted status, review corrections and regression outcomes in the cost record.
Three worked PR examples
1. Typo and test-name cleanup
Assume 20k input and 3k output on GPT-5.4 mini. Calculated cost is $0.0086. The agent should not browse the whole repository or run a full integration suite.
2. Validation refactor
Assume 100k input and 15k output on GPT-5.6 Terra. Calculated cost is $0.1425. If one failed attempt consumes the same tokens before the accepted run, total becomes $0.2850.
3. New authenticated feature
Assume 400k input and 60k output on GPT-5.6 Sol. Calculated model cost is $1.14. If the same task used Opus 4.8 under the same token assumption, it would calculate to $1.75. This does not predict which model would produce the accepted PR.
A tool-agnostic routing policy
This framework applies to Aider, Cline, Codex, OpenHands and other agent harnesses.
| PR class | Start route | Escalate when |
|---|---|---|
| Mechanical, low-risk | gpt-5.4-mini | Formatting/schema failure or unexpected reasoning need |
| Routine feature or bug | gpt-5.6-terra | Failed test diagnosis or repeated incorrect edit |
| High-risk, multi-file | gpt-5.6-sol | Start here when rework cost is material |
| Claude-specific evaluation | claude-sonnet-4-6 | Escalate to Opus only after measured failure |
For coding-tool differences, see Aider versus Cline and Claude Code versus Codex CLI.
How to cap spend without harming delivery
- Label PR size before invoking the agent.
- Set a maximum turn count and token budget.
- Exclude irrelevant directories and truncate tool output.
- Use the cheaper route for mechanical exploration.
- Escalate once, based on a documented failure condition.
- Stop after repeated identical failures.
- Attach request ids and usage records to the PR.
- Review cost per accepted PR weekly.
The broader agent API bill guide covers caching, batching and routing controls.
Monthly scenario: 200 pull requests
Assume a ten-person team produces:
- 80 small PRs on Terra at $0.0285 = $2.28
- 90 medium PRs on Terra at $0.1425 = $12.83
- 30 large PRs on Sol at $1.14 = $34.20
The calculated monthly model total is $49.31 before failures. At a 20% retry overhead, it becomes about $59.17. Infrastructure, developer time and code review are separate.
A $49 model bill is excellent only if the changes are accepted and safe. If 150 of 200 are accepted, calculated model cost per accepted PR is $0.33 before retry overhead.
What most guides get wrong
The usual mistake on How Much Does a Coding Agent API Cost per Pull Request? is to chase the lowest posted per-million-token rate while ignoring orchestration: repeated context, oversized tool output, silent retries, and premium-model defaults often dominate the invoice.
A cheaper list price does not help when every agent hop resends the same repository map or when failed tests trigger unbounded re-runs. Optimize cost per successful task—not the price of a single completion—before you renegotiate vendor contracts or swap gateways.
A realistic production scenario
A six-person platform team standardized on one premium model for How Much Does a Coding Agent API Cost per Pull Request? after a strong demo score. Early invoices looked fine; by week three, failing test loops had doubled input tokens without more accepted output.
They kept the orchestrator but tightened routing: a lower-cost catalog model handled exploration, a mid-tier model handled implementation, and premium routes required a logged failure condition. They truncated test logs and stopped attaching full diffs on every retry.
The breakthrough was reconciling LumeAPI usage records with task ids. Token list price mattered less than eliminating duplicate context across retries.
Expert take
LumeAPI is an independent third-party OpenAI-compatible gateway—not OpenAI, Anthropic, or Google. Published catalog rates can be materially below official reference pricing, but compatibility does not guarantee identical behavior for every native feature, streaming edge case, tool schema, Batch workflow, or prompt-cache semantics.
Follow this advice when you control prompts, retries, and routing and need auditable USD billing across GPT, Claude, and Gemini on one key. Do not cut over blindly if you depend on provider-native capabilities your stack has not tested end to end.
Run a bounded smoke test on your exact SDK, model id, streaming mode, and tool path; reconcile usage logs with orchestrator task ids; and keep a fallback route until error rates and latency are acceptable.
FAQ
Are coding agents billed per pull request?
Usually no. Model APIs bill tokens or another usage unit. “Per PR” is an internal allocation metric built from request logs.
Does a bigger diff always cost more?
Not always, but larger scope usually increases file reads, tool rounds and test output. Measure rather than infer from changed lines alone.
Should teams use the cheapest model for every PR?
No. Optimize accepted-PR cost. A model that requires repeated retries can be more expensive overall.
How do I reconcile the number?
Store agent task id, repository/PR id, model, request id, tokens, retries, billed USD and accepted outcome. Reconcile those fields with LumeAPI usage logs.
Sources and methodology
- LumeAPI model catalog — ids and USD rates checked July 26, 2026
- LumeAPI usage-log guide
- AI-agent API cost controls
All PR examples are labeled scenarios. No customer billing data is used.
Official references: OpenAI API pricing · Anthropic pricing · Google Gemini API pricing