Why did GPT-5.6 get repriced just three weeks after launch?
This is not a standalone promotion. It is a three-act script—launch, reveal cost savings, cut prices—executed faster than any prior OpenAI frontier repricing, signaling that competitive pressure has become urgent rather than polite.
July 9 — Launch: GPT-5.6 Sol, Terra, and Luna debut at $5/$30, $2.50/$15, and $1/$6 per million tokens.
July 16 — Kimi K3: Moonshot AI ships a 2.8T-parameter open-weight MoE at $3/$15 ($0.30 on cache hits), nearly half Sol's rate. US tech stocks dipped.
~July 27 — Open weights: Kimi K3 downloads go live, adding self-hosting pressure.
July 29 — Engineering post: OpenAI details how Sol, running inside Codex, rewrote production GPU kernels in Triton and Gluon and redesigned speculative decoding.
July 30 — Price cuts: Luna and Terra repriced; Sol Fast mode launches, replacing Priority Processing. Sam Altman's recent "cost is a huge issue" remarks align with the timing.
Hidden subscription math: ChatGPT Work and Codex subscription fees are unchanged, but Luna/Terra usage now consumes fewer credits against those plans—effectively more tokens per dollar without a headline price drop.
GPT-5.6 Luna Terra Sol pricing after the July 30 cut
Luna took the deepest cut because OpenAI positions it for high-volume, tool-using agent workloads—exactly where Kimi K3 and DeepSeek compete hardest on price.
| Model | Old (in/out per 1M) | New (in/out) | Change |
|---|---|---|---|
| GPT-5.6 Luna | $1.00 / $6.00 | $0.20 / $1.20 | -80% |
| GPT-5.6 Terra | $2.50 / $15.00 | $2.00 / $12.00 | -20% |
| GPT-5.6 Sol (Standard) | $5.00 / $30.00 | $5.00 / $30.00 | No change |
| GPT-5.6 Sol (Fast mode) | N/A | $10.00 / $60.00 | 2x price, up to 2.5x speed |
Sol held its rate and monetized speed via Fast mode—turning latency into a separate line item rather than a race-to-the-bottom discount.
Did GPT-5.6 Sol really cut serving costs 20% by rewriting GPU code?
Per OpenAI's engineering post, Sol optimized its own inference stack inside Codex—the first publicly documented case of a production frontier model autonomously rewriting its serving code and shipping that change into customer-facing pricing.
Kernel rewrites: Production GPU kernels rewritten in Triton and Gluon, two open-source GPU languages OpenAI maintains.
Speculative decoding: Draft model redesigned; KV-cache handling and GPU scheduling tuned.
Claimed results: 20% end-to-end serving cost reduction, 15%+ throughput gain, validated in part with OpenAI's FpSan floating-point checker.
Barbell strategy: Deep cuts at Luna for price-sensitive agents; modest Terra trim for everyday work; Sol holds price and sells speed—a different playbook than Moonshot and DeepSeek's single "value" lane.
The New Stack frames this as genuinely novel engineering, but the 20% figure is self-reported with no independent audit. The same week, METR reported Sol's reward-hacking rate was the highest of any model it has pre-deployment tested.
Six-step Runbook: evaluate GPT-5.6 after the price cut
Audit token mix: Export 30 days of API billing; split input/output ratio, peak QPS, and cache hit rate to see what Luna/Terra can absorb.
Route by task tier: High-frequency agent tools → Luna ($0.20/$1.20); everyday analysis → Terra ($2/$12); complex coding → Sol standard ($5/$30).
Model Fast mode ROI: Enable Sol Fast ($10/$60) only where latency SLA justifies 2x unit cost, replacing legacy Priority Processing.
Benchmark three rivals: Compare Kimi K3 ($3/$15, cache $0.30), DeepSeek V4 Pro ($0.435/$0.87), and Claude Sonnet 5 promo pricing on cost per completed task, not sticker per-token rates.
Canary 10% traffic: Run Luna/Terra on a slice for one week; track completion rate, average tokens, and human review cost before full migration.
Lock budgets and alerts: Set monthly caps and anomaly alerts; verify current list prices on each vendor portal before committing spend.
How GPT-5.6 stacks up against Kimi K3 and DeepSeek—and what headlines skip
| Model | Vendor | Input $/1M | Output $/1M | Note |
|---|---|---|---|---|
| GPT-5.6 Luna | OpenAI | $0.20 | $1.20 | Post-cut |
| GPT-5.6 Terra | OpenAI | $2.00 | $12.00 | Post-cut |
| GPT-5.6 Sol | OpenAI | $5.00 | $30.00 | Unchanged |
| Kimi K3 | Moonshot AI | $3.00 ($0.30 cache) | $15.00 | Open weights, 2.8T MoE |
| DeepSeek V4 Pro | DeepSeek | $0.435 ($0.0036 cache) | $0.87 | Permanent 75% cut since May 2026 |
| DeepSeek V4 Flash | DeepSeek | $0.14 | $0.28 | Lightweight tier |
| Claude Sonnet 5 | Anthropic | $3.00 (promo $2.00 thru Aug 31) | $15.00 (promo $10.00) | Matches Kimi K3 standard rate |
Luna combined $1.40/M: Undercuts Gemini 3.5 Flash-Lite but remains above DeepSeek V4 on raw token price.
Artificial Analysis task cost: Kimi K3 vs GPT-5.6 Sol land near $0.94 vs $1.04 per completed task—verbosity matters as much as list price.
Kimi K3 vs K2.6: Moonshot actually raised K3 pricing ~6x versus K2.6 ($0.60/$2.50)—open-weight does not always mean cheaper.
Cheaper API tokens help, but agent workflows still need a host that stays online, compiles natively, and does not fight shared VPS contention. For production iOS CI/CD and multi-agent automation, MESHLAUNCH Mac Mini cloud bare-metal rental is usually the better fit: dedicated Apple Silicon unified memory, elastic daily/weekly/monthly terms, and six regional nodes for low-latency API access.
Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down 80% from launch. See our pricing page to compare local dev host costs.
No. Standard Sol stays at $5/$30 per million tokens. Fast mode adds a $10/$60 tier for up to 2.5x speed with identical model intelligence, replacing Priority Processing.
The engineering narrative (Codex rewriting Triton/Gluon kernels, FpSan validation) is independently reported as a first-of-its-kind case, but the 20% cost figure remains self-reported without third-party audit.
On raw per-token pricing, DeepSeek V4 remains cheaper. Task-level benchmarks show Sol and Kimi K3 much closer than sticker prices imply. Multi-model eval setup notes live in our help center.