GPT-6 Luna is cheaper for coding agents: both list at $0.10 per million input and $0.50 per million output tokens, but Claude Haiku 5.5 charges 5x once a prompt passes 100,000 tokens, while Luna's smaller surcharge starts at 272,000. Priced on a month of real coding-agent traffic from our gateway, Haiku 5.5 cost 3.1x as much as Luna.
Below 100,000 input tokens the two models cost exactly the same. Every difference in the bill comes from larger prompts.
| Prompt size | Haiku 5.5 | Luna |
|---|---|---|
| Up to 100K | $0.10 · $0.01 · $0.50 | $0.10 · $0.01 · $0.50 |
| 100K to 272K | $0.50 · $0.05 · $2.50 | $0.10 · $0.01 · $0.50 |
| Over 272K | $0.50 · $0.05 · $2.50 | $0.20 · $0.02 · $0.75 |
Per 1M tokens: input · cached input · output. Standard API rates, checked October 13, 2026. A request over the line pays the higher rates on all of its tokens.
Assumptions from our measured averages: 87% of input tokens are cache reads, and a request returns about 700 output tokens. A 150K prompt on Haiku 5.5 is then (19,500 fresh × $0.50 + 130,500 cached × $0.05 + 700 output × $2.50) per million tokens.
| Prompt size | Haiku 5.5 | Luna |
|---|---|---|
| 90K | $0.0023 | $0.0023 |
| 150K | $0.0180 | $0.0036 |
| 300K | $0.0343 | $0.0135 |
The gap widens over a session, because the agent's context grows with every step. Each logged request was priced at both list prices using its own token counts. A session is one key with no pause longer than 10 minutes.
| Measured cost | Haiku 5.5 | Luna |
|---|---|---|
| Per 1,000 requests | $13.07 | $4.16 |
| Median session (16 requests) | $0.096 | $0.046 |
| Average session (67 requests) | $0.88 | $0.28 |
1,177 sessions. Without the five heaviest users the ratio is 2.9x instead of 3.1x.
A matching price card does not mean a matching bill. Coding agents resend the whole conversation on every step, so prompts grow as a session runs and cross 100K sooner than most people expect. Token counts also differ between tokenizers: ours come from the models that actually served the traffic, and the same text can count somewhat differently on Claude or GPT. Test both models on a task from your own repository before switching.
Claude Haiku 5.5 and GPT-6 Luna are both in the Standard Compute catalog. Select either one in Dashboard → Models, or leave smart routing on and let the router choose per request. Standard Compute starts at $19/month with a $20 monthly compute budget. Requests stop when that budget is exhausted until renewal or a plan change; there are no automatic overage charges. Optional pacing can spread usage across the month.
Not at list price. Up to 100,000 input tokens the rates are identical. Between 100K and 272K, Haiku 5.5 charges its higher band while Luna stays at base rates. Over 272K both charge more, and Haiku 5.5's rates are still higher on input, cache and output.
Yes. Anthropic counts every input token, including cache reads and cache writes, and a request over the line pays the higher rates even when most of its prompt is a cache hit. OpenAI applies Luna's long-context rates to the full request once input exceeds 272K tokens.
Claude Haiku 5.5 accepts 1M tokens and GPT-6 Luna 1,050,000, each with up to 128K output tokens, according to the vendors' model pages. Both windows are far larger than either price threshold, so for coding agents the price band matters long before the window does.
Start a new session for each task, let the agent compact its history earlier, and hand self-contained subtasks to subagents with a fresh context. Our median request sat right at the line, so small habits move a large share of traffic across it.
No. Both vendors halve their rates for batch jobs and keep the same thresholds, so the ratio holds. Interactive coding agents usually cannot wait for batch results anyway.
From Standard Compute's gateway logs for Sep 13 to Oct 12, 2026, which record token counts and the client application for each request, never prompt text. We counted successful requests from clients that identify as coding agents (OpenCode, Kilo, Roo and GitHub Copilot Chat): 79,292 requests from 75 users. Only OpenCode had 50 or more users, so the agents are pooled, and we publish aggregates only. The median request carried 99,545 input tokens, p90 279,944 and p99 533,300, counted by the model that served it. Cache writes, under 1% of input tokens, are priced at each model's write rate, which is 1.25x its input rate.