GPT-5.6 Luna costs a fifteenth of Claude Sonnet 4.6. If you are still on Sonnet 4.6, it is not the only exit to price.
Luna lists at $0.20 / $1.20 against Sonnet 4.6's $3 / $15 and runs nearly three times faster. Why Claude Sonnet 5 belongs in the same decision.
GTM Architect & Growth Operator · Now · 16 September 2026
TL;DR · Key insights
- GPT-5.6 Luna lists at $0.20 / $1.20 per million tokens and Claude Sonnet 4.6 at $3 / $15: 15 times cheaper on input and 12.5 times on output. Above 272K tokens Luna rises to $0.40 / $1.80 and is still several times cheaper.
- Luna is also nearly three times as fast on Artificial Analysis's measurements, about 115 output tokens per second against Sonnet 4.6's 42.
- On Intelligence Index v4.3, Luna at max effort scores 37.50. The only Sonnet 4.6 figure is its non-reasoning variant at 24.69, so read the gap as direction, not distance.
- Anthropic lists Sonnet 4.6 as a legacy model. If you are leaving it, Claude Sonnet 5 at $2 / $10 costs less than Sonnet 4.6 and scores within a point of Luna, so it belongs in the same decision.
Searches for GPT-5.6 Luna against Claude Sonnet 4.6 mostly come from one situation: a team still running Sonnet 4.6 and looking at a price list that no longer makes sense.
On the numbers the comparison is lopsided. The more useful question is what else belongs next to it.
| Info | GPT-5.6 Luna | Claude Sonnet 4.6 |
|---|---|---|
| Input / output per 1M | $0.20 / $1.20 | $3 / $15 |
| Cached input per 1M | $0.02 | $0.30 |
| Above 272K input tokens, in / out | $0.40 / $1.80 | $3 / $15 |
| Intelligence Index v4.3 | 37.50 (max) | 24.69 (non-reasoning) |
| Output speed, tokens/s | about 115 | about 42 |
| Context window | 1,050,000 | 1M |
| Status | Current | Legacy, still available |
Prices from OpenAI's GPT-5.6 Luna model page and Anthropic's pricing page; Sonnet 4.6's status from Anthropic's models and deprecations pages. Index to two decimals and speed from Artificial Analysis, read on 16 September 2026. The two index figures are different variants, so read them as direction.
Fifteen times cheaper, and the long-context rule does not change that
Luna lists at $0.20 input and $1.20 output per million tokens. Sonnet 4.6 lists at $3 and $15. On every rate, including cached input, Luna is 12.5 to 15 times cheaper.
OpenAI’s long-context surcharge narrows the gap without closing it. Above 272K input tokens Luna’s whole request moves to $0.40 / $1.80, which is still 7.5 times cheaper than Sonnet 4.6 on input and more than 8 times on output. Anthropic bills Sonnet 4.6 at one rate across its 1M window, and at these prices that one rate does not rescue it.
Nearly three times as fast
Artificial Analysis measures Luna at about 115 output tokens per second and Sonnet 4.6 at about 42. In time, a 1,000-token answer takes about nine seconds on Luna and about 24 on Sonnet 4.6.
The index row needs a caveat
On Intelligence Index v4.3, Luna at max effort scores 37.50. Sonnet 4.6 has no reasoning variant on this version of the index, so the figure available is its non-reasoning variant at 24.69.
That is not a like-for-like pair. I read it as direction rather than distance, and a gap of nearly 13 points is wide enough that I would trust the direction.
On price and speed, Luna wins by a wide margin. On capability the published numbers point the same way, without a clean comparison behind them.
The comparison you should actually make
Anthropic lists Sonnet 4.6 as a legacy model. It still works. Its deprecations page commits to retirement not sooner than 17 February 2027, so nothing forces a move this year. If you are leaving anyway, there are two exits, and they answer different questions.
- GPT-5.6 Luna answers “what is the cheapest model that does this well enough?” It costs a fifteenth of Sonnet 4.6 on input.
- Claude Sonnet 5 answers “what replaces Sonnet 4.6 without leaving Anthropic?” It costs $2 / $10, less than Sonnet 4.6, scores 38.36 on the same index, within a point of Luna, and generates at about 75 tokens per second.
Sonnet 5 carries one cost of its own. Its newer tokenizer produces about 30% more tokens for the same text, which I covered in Sonnet 4.6 against Sonnet 5. Against Luna it is still ten times the input price, the trade I went through in Luna against Sonnet 5.
Which one I would use
| The work | Model | Why |
|---|---|---|
| Classification, extraction, routing at volume | GPT-5.6 Luna | A fifteenth of the input price and nearly three times the speed. |
| Latency-sensitive assistants | GPT-5.6 Luna | About 115 tokens per second against 42. |
| Work built around Claude prompts and tools | Claude Sonnet 5 | Cheaper than Sonnet 4.6, within a point of Luna on the index, same platform. |
| Staying on Sonnet 4.6 | Neither | A legacy model that now costs more than both alternatives. |
The real choice is Luna or Sonnet 5. Sonnet 4.6 loses to both on price.
For the cheaper Claude tier, see Haiku 4.5 against Luna.
What would change my mind
A reasoning-variant score for Sonnet 4.6 on v4.3. It would turn the index row from direction into distance.
A price cut on Sonnet 4.6. At $3 / $15 it sits above Sonnet 5 on both sides, which is unusual for an older model and could change.
An earlier retirement date. Anthropic’s current commitment gives Sonnet 4.6 until at least February 2027, which is time to test properly rather than move in a hurry.
Questions people asked
Is GPT-5.6 Luna better than Claude Sonnet 4.6?
On the numbers available, yes, though not on a like-for-like index row. Artificial Analysis’s Intelligence Index v4.3 scores Luna at max effort at 37.50, and the only Sonnet 4.6 score is its non-reasoning variant at 24.69. Luna is also nearly three times as fast, at about 115 output tokens per second against 42, and far cheaper. Whether your prompts and tools behave the same on an OpenAI model is the part to test.
How much cheaper is GPT-5.6 Luna than Claude Sonnet 4.6?
Luna costs $0.20 per million input tokens and $1.20 output against Sonnet 4.6’s $3 and $15: 15 times cheaper on input and 12.5 times on output. Cached input is $0.02 against $0.30. Above 272K input tokens OpenAI bills Luna’s whole request at $0.40 input and $1.80 output, still several times below Sonnet 4.6, which keeps one rate across its 1M window.
Should I switch from Claude Sonnet 4.6 to GPT-5.6 Luna?
For high-volume, well-defined work such as classification, extraction and routing, it is worth testing, because the price and speed gaps are large. For work built around Claude’s tools and prompts, price Claude Sonnet 5 too: it costs $2 / $10, less than Sonnet 4.6, scores within a point of Luna on v4.3 and keeps you on Anthropic’s platform. Anthropic lists Sonnet 4.6 as a legacy model, with retirement not sooner than 17 February 2027.
What about Claude Sonnet 5 instead?
Sonnet 5 is the other exit from Sonnet 4.6. It lists at $2 / $10, scores 38.36 on the Intelligence Index v4.3 against Luna’s 37.50, and runs at about 75 output tokens per second against Sonnet 4.6’s 42. It is still ten times Luna’s input price, and its newer tokenizer turns the same text into about 30% more tokens than Sonnet 4.6 does. Choose it when staying on Anthropic matters more than the price gap.
If you are planning a move off Sonnet 4.6 and want the numbers run on your own workload, the contact page is the fastest route to me.
