Blog · 2026-07-17 · Vynaris Team
GPT-5.6 Sol vs Terra vs Luna, priced against Fable 5 and Opus 4.8: cost per coding task
GPT-5.6's three tiers priced per coding task: gpt-5.6-luna $200/1k to gpt-5.6-sol $1,001/1k. Opus 4.8 undercuts Sol on the list price, but its tokenizer flips the verdict. Prices verified 2026-07-17.
OpenAI shipped GPT-5.6 in three tiers, and on a plan-edit-verify coding task the tier you pick swings the bill about 5x before you change a line of code: gpt-5.6-luna runs $200 per 1,000 tasks, gpt-5.6-sol $1,001. The counterintuitive part: Claude Opus 4.8 looks cheaper than Sol on the price list, until its tokenizer flips the verdict. Prices verified 2026-07-17.
TL;DR
- Same coding task, three GPT-5.6 tiers: $200.22 per 1,000 tasks on gpt-5.6-luna, $500.56 on gpt-5.6-terra, $1,001.12 on gpt-5.6-sol. Terra is exactly 50% of Sol and Luna is 80% less, on both input and output, so the tier is a straight dial on your bill.
- gpt-5.6-sol and Claude Opus 4.8 share the same $5 input price, but Opus output is $25 against Sol's $30. On raw per-token math Opus 4.8 lands 5.3% cheaper than Sol. Apply Anthropic's stated ~30% tokenizer inflation and Opus 4.8 becomes 23% more expensive than Sol on the same task. The tokenizer is the whole verdict.
- Route the mechanical steps down a tier and keep planning on Sol: a plan-on-Sol, edit-and-verify-on-Terra split costs 31% less than all-Sol; splitting to Luna cuts 50%, with no change to plan quality.
What we computed, and why
A chat request is one round trip. A coding agent is a loop: it reads an issue, loads files, plans, edits, runs tests, reads the failures, patches, and repeats. Each pass is a separate model call, and each call re-sends the system prompt, the tool schemas, and the code context. That structure is why agent inference costs behave nothing like the per-message math carried over from chatbots, and why a per-token price list does not tell you what a task costs.
So we built the cost per coding task, not the cost per million tokens. Every list page (pricepertoken, the aggregators) will tell you Sol is $5/$30. None of them tell you what that means for a plan-edit-verify loop with a reasoning budget. We priced the three GPT-5.6 tiers against Claude Fable 5 and Claude Opus 4.8 on one representative task, with every assumption in a table you can edit. All arithmetic is in a script; nothing here was computed by hand.
The assumptions (edit these)
Assumption Value Note
------------------------- ---------------------------- ---------------------------------------------------------------------------
Stable prefix per call 15,000 tokens system prompt + tool schemas + loaded code, re-sent each call unless cached
Model calls per task 8 plan (1) + edit (3) + verify (2) + 2 back-and-forth
Fresh input per call 2,000 tokens new tool results and appended messages, never cacheable
Visible output per call 838 tokens diffs + tool calls, averaged
Reasoning tokens per call 500 (low) / 8,000 (high) billed at the output rate, invisible in the response
Total input per task 136,000 tokens 8 x (15,000 + 2,000)
Total output per task 10,704 (low) / 70,704 (high) 8 x (838 + reasoning)The reasoning-token budget is the axis that matters here. Reasoning tokens never appear in the response, but they bill at the full output token rate. At the low setting the task is input-heavy (12.7:1 input to output); at the high setting the loop does deep planning and the ratio collapses to 1.9:1, which hands the verdict to whichever model has the cheaper output price. We show both.
Results: cost per task, raw per-token prices
Prices verified 2026-07-17 from the OpenAI and Anthropic pricing pages (sources below). Standard first-party rates, no caching applied yet, low-reasoning shape.
Model Input $/1M Output $/1M Cost/1,000 tasks
--------------------------------------------------------- ---------- ----------- ----------------
gpt-5.6-luna $1.00 $6.00 $200.22
[gpt-5.6-terra](https://vynaris.com/models#gpt-5-6-terra) $2.50 $15.00 $500.56
Claude Opus 4.8 $5.00 $25.00 $947.60
gpt-5.6-sol $5.00 $30.00 $1,001.12
Claude Fable 5 $10.00 $50.00 $1,895.20
Two things fall out. First, the GPT-5.6 tiers are a clean dial: Terra bills exactly half of Sol and Luna exactly a fifth, because each tier scales input and output by the same factor. Pick the tier and you have set the bill. Second, Sol is not the cheapest way to buy a $5-input frontier coding model here. Opus 4.8, at the same $5 input and a lower $25 output, comes in 5.3% under Sol on this shape. Which brings us to the catch.
The Sol vs Opus 4.8 flip nobody prints on the pricing page
gpt-5.6-sol and Claude Opus 4.8 have identical $5 input pricing. The only per-token difference is output: $30 on Sol, $25 on Opus. So on any output-heavy shape Opus should win, and the raw math says it does, by 5.3% at low reasoning and 12.6% at high reasoning.
Then you read Anthropic's own footnote. Opus 4.7 and later, plus Fable 5, use a newer tokenizer that "produces approximately 30% more tokens for the same text." GPT-5.6 uses OpenAI's tokenizer. That means a like-for-like comparison, same code and same English prompt, bills the Anthropic models for roughly 30% more tokens than the token counts above assume. Apply it:
Reasoning shape gpt-5.6-sol Opus 4.8 raw Opus 4.8 +30% tokenizer
--------------- ----------- ------------------------- -----------------------
Low $1,001.12 $947.60 (5.3% cheaper) $1,231.88 (23% dearer)
High $2,801.12 $2,447.60 (12.6% cheaper) $3,181.88 (14% dearer)The tokenizer, not the sticker, decides this matchup. On the price list Opus 4.8 undercuts Sol; on the same task, adjusted for how many tokens the same text becomes, Sol is cheaper by 14% to 23%. We wrote up the mechanism in full in the tokenizer post; the 30% is Anthropic's stated average and varies with content, so measure it on your own prompts with the count_tokens API before trusting it to the dollar. But do not compare token prices across tokenizer families as if a token means the same thing on both.
This is the moment to plug in your own numbers: run the per-task math in the calculator with your call count, context size, and reasoning budget before you commit to a model.
Where the reasoning budget changes the answer
At the low-reasoning setting the task is input-bound, so the ranking tracks input price. Crank reasoning to 8,000 tokens per call and the picture tilts toward output price, because reasoning tokens bill as output:
Model Cost/1,000 (low) Cost/1,000 (high) Change
--------------------- ---------------- ----------------- ------
gpt-5.6-luna $200.22 $560.22 +180%
gpt-5.6-terra $500.56 $1,400.56 +180%
Claude Opus 4.8 (raw) $947.60 $2,447.60 +158%
gpt-5.6-sol $1,001.12 $2,801.12 +180%
Claude Fable 5 (raw) $1,895.20 $4,895.20 +158%A reasoning-heavy loop nearly triples every bill, and the models with the widest output-price gap (Sol at $30, Fable at $50) absorb the most damage. If your agent runs long chains of thought, the reasoning budget is a bigger lever on the invoice than the tier choice. We took the reasoning-token line item apart in the reasoning-token tax post; the short version is that capping the budget is often the cheapest optimization on the board.
What it means for routing
Right-sizing an agent is not one decision. The plan step and the verify step have different quality bars, and the three GPT-5.6 tiers give you a clean way to split them.
- Route the plumbing, keep the planner. The verify step reads a test failure and applies a mechanical patch; it rarely needs the top tier. Keep planning on Sol and move edits and verification to Terra or Luna. On our task, plan-on-Sol with edit-and-verify-on-Terra costs $688.27 per 1,000 (31% less than all-Sol); splitting to Luna costs $500.56 (50% less). Plan quality is untouched because planning still runs on Sol.
- The tier is a linear dial, so size it to the step. Because Terra is exactly half of Sol and Luna a fifth, you can reason about the split in your head: every call you move from Sol to Luna removes 80% of that call's cost.
- For cross-vendor calls, price the tokenizer, not the sticker. Sol versus Opus 4.8 is a coin flip on the price list and a clear Sol win once you count tokens. Any routing table that compares OpenAI and Anthropic on headline $/MTok is comparing two different units.
An OpenAI-compatible gateway makes the split testable without rewriting provider glue: point one base URL at a router, send verify-step calls to Luna, keep planning on Sol, and compare the invoice. That is the model routing case in one sentence.
Where this model is wrong
- Retries are not in the token counts. A cheaper tier fails the verify loop more often. If Luna needs three extra calls where Sol needs zero, the per-task gap narrows. Route the cheap tier onto steps where a wrong answer is cheap to catch, not onto planning, where a bad plan wastes every downstream call.
- Token counts are illustrative, not measured from your traffic. We used a public, editable assumption set. Your prefix might be 40k tokens, not 15k, and your reasoning budget might sit between our two settings. Re-run the script with your trace.
- The tokenizer adjustment is an average. The stated 30% varies by content and workload shape. It is the load-bearing number in the Sol-vs-Opus verdict, so measure it before you route on it.
- Caching is off in every number here. Turning on prompt caching cuts the stable-prefix cost sharply on all five models and would reorder some of the close calls. We priced the uncached case so the tier and tokenizer effects are visible on their own.
When does the top tier win outright? When one bad plan cascades. On a task where a flawed plan burns all eight calls and then a human's afternoon, the Sol run is cheaper than three failed Luna runs plus the cleanup. Frontier pricing buys fewer retries, and on the planning step that is often the right trade.
FAQ
How much does a coding task cost on GPT-5.6? On this model (8 calls, 136k input and ~10.7k output tokens, low reasoning), one task costs $0.20 on gpt-5.6-luna, $0.50 on gpt-5.6-terra, and $1.00 on gpt-5.6-sol at standard rates verified 2026-07-17. A reasoning-heavy shape roughly triples each figure.
Is gpt-5.6-sol cheaper than Claude Opus 4.8? On the raw price list, no: Opus 4.8 shares Sol's $5 input and undercuts its output ($25 vs $30), landing about 5% cheaper per task. Once you apply Anthropic's stated ~30% tokenizer inflation, Sol is cheaper by 14% to 23% on the same text. The tokenizer decides it.
What is the difference between Sol, Terra and Luna? Price, on a fixed ratio. Terra bills exactly half of Sol on both input and output; Luna bills a fifth. On a coding task that makes Terra 50% cheaper than Sol and Luna 80% cheaper, before any quality difference.
Should I just use the cheapest tier? Only for steps with a low quality bar. Keep planning on Sol and route mechanical edits and verification to Terra or Luna; that captured 31% to 50% of the cost here without touching plan quality. A weaker tier on the planning step can cost more via retries than it saves per call.
Why does the reasoning budget matter so much? Reasoning tokens bill at the output rate but never appear in the response. At 8,000 reasoning tokens per call the task cost roughly triples, and the models with the highest output price take the biggest hit. Capping the reasoning budget is frequently a larger saving than switching tiers.
Sources
- OpenAI API pricing (GPT-5.6 Sol/Terra/Luna input, output, cached input and cache-write rates), captured 2026-07-17: https://developers.openai.com/api/docs/pricing
- Anthropic pricing (Claude Opus 4.8 and Fable 5 rates, the ~30% tokenizer note, prompt-caching multipliers), captured 2026-07-17: https://platform.claude.com/docs/en/docs/about-claude/pricing
- Cost model script and per-model math: reproducible from the assumptions table above.
Prices change. We re-verify every figure in this post monthly and stamp updates. Numbers here are current as of 2026-07-17.
Want the same math on your own workload? Vynaris is an OpenAI-compatible gateway that routes each request to the cheapest right-sized model and shows the per-request cost. One base URL swap. Get an API key at vynaris.com.