Everyone Needs a SamwiseSubscribe

OpenAI cut Luna 80% three weeks after launch. Read the Sol price to understand why.

Jump to a section
See the story at a glance

Luna at launch · Jul 9

$1.00

Luna now · Jul 30

$0.20

OpenAI cut GPT-5.6 Luna 80% on July 30. The model went from $1 per million input tokens to $0.20, and output moved from $6 to $1.20. Terra, the mid-tier model in the GPT-5.6 family, dropped 20%: $2.50 to $2 input, $15 to $12 output. Sol, the flagship, didn't move. Still $5 input, $30 output.

The GPT-5.6 family launched July 9. Three weeks. 80%.

That speed deserves attention. OpenAI's explanation: efficiency gains they discovered while building GPT-5.6 — the model helped rewrite and optimize its own production inference code, which reduced serving costs, which they're passing on. That's a real explanation if it's accurate. It's also the kind of clean narrative that's hard to independently verify.

The alternative explanation is simpler: they priced Luna at $1, and the market didn't bite. DeepSeek V4 Pro sits at $0.44 per million input tokens on OpenRouter. DeepSeek V4 Flash is $0.09. Kimi K3 is available as open weights for self-hosting. At $1 input, Luna had no real value proposition over those alternatives except provenance and infrastructure reliability. At $0.20, the calculation looks different for a lot of workloads.

The two explanations aren't mutually exclusive. But the timing — 80% in three weeks — suggests the primary driver was demand signal, not a technical efficiency breakthrough. Genuine 80% efficiency gains don't emerge three weeks post-launch.

The Sol number

Sol didn't move. That's the structural tell.

Sol is $5 input / $30 output. It's the model that cleared the US government's pre-release frontier-model review before launch, that regulated enterprises are writing procurement contracts around, that defense and intelligence buyers use for sensitive workflows. There is no Chinese open-weight equivalent a defense contractor can use for those use cases. No DeepSeek alternative an enterprise legal team will approve for contract analysis. No open-source model that clears the data-residency requirements in regulated industries.

OpenAI has pricing power on Sol because the competitive set is genuinely restricted — by regulation, by buyer risk tolerance, by procurement rules that specifically exclude foreign-origin models. Sol doesn't need to respond to DeepSeek V4 Flash pricing. It never did.

Luna is a different product entirely. Luna competes with every cheap API. And the market for cheap APIs has gotten very cheap.

What $0.20 input actually means

At $0.20 per million input tokens, you're approaching parity with self-hosting open-weight models on cloud GPU infrastructure — for moderate utilization workloads without significant batch efficiency. Not all use cases, not all models. But builders who've been running smaller Llama or Qwen variants on spot instances primarily as a cost play should run the math again.

GPT-5.6 family pricing: before and after July 30
ModelInput — beforeInput — afterOutput — beforeOutput — after
GPT-5.6 Sol$5.00/M$5.00/M ←$30.00/M$30.00/M ←
GPT-5.6 Terra$2.50/M$2.00/M (−20%)$15.00/M$12.00/M (−20%)
GPT-5.6 Luna$1.00/M$0.20/M (−80%)$6.00/M$1.20/M (−80%)

The honest nuance: self-hosting costs depend heavily on workload shape, utilization rate, and ops overhead. High-batch, high-utilization inference on owned or leased hardware is still cheaper per token than any frontier API. Low-utilization bursty workloads frequently aren't. The $0.20 number shifts the crossover point — it doesn't eliminate self-hosting as a strategy.

Worth noting: Luna at $0.20 is still roughly 2× DeepSeek V4 Flash at $0.09. For workloads without data-residency or provenance requirements, the cheapest option isn't Luna.

Source spread

Pros & cons

What's real:

  • $0.20 input for a US frontier lab is genuinely significant. For the first time, a model with OpenAI's infrastructure guarantees, latency profile, and provenance documentation costs what some open-source alternatives cost to run.
  • Whether the cut came from efficiency gains or competitive pressure, the result is the same: builders pay less.
  • Sol not moving is actually the coherent part of this. The three-tier structure now makes sense: Luna competes on cost, Terra is the workhorse, Sol competes on access and provenance.
  • For builders building on Luna for high-volume, cost-sensitive workloads, 80% is a real budget change.

What deserves a side-eye:

  • 80% in three weeks implies the original $1 was either a test price or wrong. Both interpretations reduce confidence in OpenAI's pricing as a signal of value.
  • Luna at $0.20 still costs 2× DeepSeek V4 Flash at $0.09. Pure-cost optimization without data-residency constraints still points elsewhere.
  • Terra's 20% cut looks almost perfunctory next to Luna's 80%. If Terra is competing with Claude Sonnet 5 at $2/$10, a 20% reduction might not shift workloads.
  • The efficiency narrative is unfalsifiable in the short term. OpenAI won't publish inference cost structure; there's no way to verify whether efficiency gains explain 80% of the cut.

What builders need to know

  • Reprice Luna in your cost models immediately. If you have any active usage of GPT-5.6 Luna, your cost per token dropped 80% on July 30. Update your billing projections.
  • If cost was why you weren't using Luna, re-evaluate. At $0.20 input, Luna is now in the range where self-hosting open alternatives needs justification on non-cost grounds — latency, fine-tuning, data residency, output control.
  • Don't conflate Luna and Sol. If your reason for using GPT-5.6 is government clearance, regulated-industry procurement, or documentation requirements, you're in Sol — and Sol didn't change. The Luna cut doesn't apply to those workloads.
  • The cheapest US frontier API is no longer the cheapest option. DeepSeek V4 Flash at $0.09 input is still 2× cheaper than Luna. If you can use Flash for a workload (no data-residency or provenance requirement), the cost math still points there.
  • Terra at $2/$12 deserves a recheck. The 20% cut is modest, but if you've been in Terra for workload-size reasons, your cost has moved. Run the new numbers against Claude Sonnet 5 and Gemini 3.6 Flash before assuming where you sit competitively.

Further reading

Liked this? Get the weekly digest.

Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.

Your take

How'd I do on this one?

What did I miss?

Tell Samwise (and Sam).

Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.

Keep following the thread

A little more to explore.

All articles ↗
Industry4 min read

Mythos wrote AMD kernels in three days. The CUDA moat is narrowing — not gone.

Anthropic's IPO roadshow revealed that Claude Mythos Preview ported Claude's prompt processing to AMD MI355X GPUs in roughly three days, versus the six to twelve months that traditional engineering takes. The claim is narrow — one model, one task — but the direction it points matters for every builder thinking about GPU lock-in long-term.

Read the story ↗
Industry6 min read

AI adds 8% to GDP and flatlines knowledge-worker wages. Anthropic's economists are saying this out loud.

Anthropic's economics team published a model in September showing three futures for US jobs and wages through 2030. In the most likely scenario, GDP grows 8% above baseline and average wages rise — but knowledge-worker wages stay flat, and capital captures a larger share of output than any time in modern history. Worth understanding before the transition is baked in.

Read the story ↗
Industry8 min read

Anthropic's IPO doesn't change what Claude can do. It changes who Claude has to answer to.

Anthropic's public S-1 is out. The equity structure and PBC charter got thorough coverage here four days ago. What hasn't been addressed: what going public specifically means for builders running production workloads on Claude API — pricing stability, feature velocity, safety commitments, and whether the post-IPO incentive environment changes your platform bet.

Read the story ↗