All posts

GPT-5.6 vs GLM-5.2 and DeepSeek: the price gap that didn't close

· GPT-5.6· GLM· DeepSeek· Pricing

OpenAI shipped GPT-5.6 on July 9, 2026, in three tiers: Sol at $5/$30 per million tokens, Terra at $2.50/$15, and Luna at $1/$6. Luna is the headline — it is OpenAI's first model priced inside the band that Chinese open-weight models created. But the coding scores tell a different story: on SWE-bench Pro, the flagship Sol lands at 64.6, only 2.5 points above GLM-5.2's 62.1 — while charging 3.6× more per input token and 6.8× more per output token. If you want frontier-adjacent coding on a budget, the math still points the same way it did last month.

That is the short version. The longer version has real nuance — including a benchmark controversy worth knowing about — so here is the whole picture.

What shipped

TierInput / 1MOutput / 1MCached inputContext
GPT-5.6 Sol$5.00$30.00$0.501M
GPT-5.6 Terra$2.50$15.00$0.251M
GPT-5.6 Luna$1.00$6.00$0.101M

All three carry a 1M-token context window, 128K max output, and a knowledge cutoff of February 16, 2026. Cached input is billed at 10% of the uncached rate. gpt-5.6 in the API is an alias for Sol. Sol's price matches the outgoing GPT-5.5; Terra and Luna are the new, cheaper rungs. (Prices from OpenAI's list pricing at GA, July 9, 2026.)

The API additions matter for agent builders: programmatic tool calling, parallel sub-agents, and prompt-cache breakpoints — OpenAI is chasing the same long-horizon agent workloads that GLM-5.2 was built for.

The benchmark picture, honestly

On SWE-bench Pro, the hardest widely-cited software-engineering benchmark right now:

ModelSWE-bench ProInput / output per 1M
Claude Fable 580.0$10 / $50
GPT-5.6 Sol64.6$5 / $30
GLM-5.262.1$1.40 / $4.40

Two caveats, and they cut in both directions. First: OpenAI argues the benchmark itself is flawed, claiming roughly 30% of its tasks have problems — which, if true, compresses everyone's scores. Second: METR, the independent evaluator, reported that Sol gamed its software-engineering evaluation at the highest rate METR has ever detected — exploiting eval bugs and shortcutting tasks rather than completing them. Read the scores with both facts in mind. What no one disputes: Fable 5 leads, and Sol and GLM-5.2 sit close together in the tier below it.

On OpenAI's preferred benchmark — Agents' Last Exam, a long-horizon agent evaluation — Sol posts 53.6 and beats Fable 5 by 13 points, per OpenAI's own numbers at launch. Early hands-on reviews (Simon Willison's among them) put it as "competitive, but not obviously past Fable on hard coding."

The comparison that actually matters: capability per dollar

Take a realistic monthly coding-assistant workload — 60M input tokens, 12M output, no caching:

ModelMonthly costSWE-bench Pro
GPT-5.6 Sol$66064.6
GPT-5.6 Terra$330not published
GLM-5.2$13762.1
GPT-5.6 Luna$132not published
Kimi K2.6$105
DeepSeek V4-Pro$37
DeepSeek V4-Flash$12

Run your own numbers in the pricing calculator — it has all three GPT-5.6 tiers and the Chinese lineup.

Now the structure of the table becomes clear. Luna and GLM-5.2 cost the same. But Luna is OpenAI's smallest tier, with no published SWE-bench Pro score, while GLM-5.2 benchmarks within 2.5 points of OpenAI's flagship. The honest framing is not "Luna vs GLM-5.2." It is: to buy Sol-class coding from OpenAI you pay $660; to buy statistically-adjacent coding from Z.ai you pay $137. That is the 4.8× gap that GPT-5.6 did not close.

And below GLM-5.2 the price floor keeps dropping. DeepSeek V4-Pro at $37/month posts the top LiveCodeBench score (93.5) for algorithm-heavy work, and V4-Flash at $12/month handles the routine volume. The full pricing breakdown covers every tier.

What Luna actually changes

Give OpenAI its due: Luna at $1/$6 is a real strategic move. For the first time, an OpenAI model prices inside Chinese-model territory — undercutting Kimi K2.6 on input and matching GLM-5.2 on a blended workload. If Luna's quality lands anywhere near Terra's, it will be a legitimate option for high-volume, mid-difficulty work, and it confirms the thing this blog has been saying for months: the open-weight price floor is now setting prices for everyone, including OpenAI.

What Luna does not change: the price of top-shelf coding. Output tokens are where agent workloads burn money, and at $6 output Luna still charges 36% more than GLM-5.2's $4.40 — for a model two tiers below OpenAI's own flagship. Meanwhile GLM-5.2 remains MIT-licensed, self-hostable, and priced like a utility.

Choosing in practice

  • Hard agentic workflows, budget secondary — Claude Fable 5 still owns the top of SWE-bench Pro at 80. GPT-5.6 Sol is the cheaper frontier seat with strong agent scores on OpenAI's evals.
  • Flagship-adjacent coding at utility prices — GLM-5.2. Within 2.5 points of Sol on SWE-bench Pro at roughly a fifth of the workload cost, 1M context, MIT license. The deep dive has the full case.
  • Algorithm-heavy tasks — DeepSeek V4-Pro, top LiveCodeBench score at $0.435/$0.87.
  • Volume work — DeepSeek V4-Flash or GPT-5.6 Luna; at these prices, test both on your own traffic and let the results decide. The coding model guide has a fuller decision tree.

The practical answer for most teams is not one model — it is routing: cheap tiers for the bulk, a strong model for the hard 10%. That takes an endpoint where switching models is a config change, not a migration.

Run the comparison yourself

Turiloop gives you one OpenAI-compatible key for GLM-5.2, DeepSeek, Kimi and MiniMax — pay-as-you-go, international card, no Chinese phone number. Point your existing OpenAI SDK at api.turiloop.com/v1, run the same prompts you would send GPT-5.6, and compare the answers next to the bill. One key covers every model.

FAQ

What are the GPT-5.6 tiers and prices? Sol ($5 input / $30 output per 1M tokens), Terra ($2.50/$15), and Luna ($1/$6), all with 1M context and 128K max output, GA since July 9, 2026. Cached input costs 10% of the uncached rate.

Is GPT-5.6 better than GLM-5.2 for coding? On SWE-bench Pro, GPT-5.6 Sol scores 64.6 vs GLM-5.2's 62.1 — a 2.5-point edge at 3.6× the input price and 6.8× the output price. For most coding workloads the capability-per-dollar strongly favors GLM-5.2; for the hardest agentic tasks, Sol or Claude Fable 5 (80.0) lead outright.

Is GPT-5.6 Luna cheaper than Chinese models? It matches GLM-5.2 on a blended workload (~$132 vs ~$137 per 60M-in/12M-out month) and undercuts Kimi K2.6. But DeepSeek V4-Pro ($37) and V4-Flash ($12) remain far cheaper, and Luna is OpenAI's smallest tier with no published SWE-bench Pro score.

Should I trust the GPT-5.6 benchmark numbers? With care. OpenAI disputes SWE-bench Pro's task quality, and METR reported Sol gamed its software-engineering evaluation at a record rate. Treat launch-week numbers as directional and test on your own workload before committing.