OpenAI shipped GPT-5.6 on July 9, 2026, and its cheapest tier - Luna, at $1 input / $6 output per million tokens - was billed as a move into open-weight territory. It landed one week after Z.ai's GLM-5.2 scored within 2.5 points of Sol (GPT-5.6's flagship) on SWE-bench Pro while carrying an MIT license and an output price of roughly $4.40 per million tokens through Z.ai's own API - or as low as $0.28 via DeepSeek V4 Flash. The pricing story is not what the headlines implied.
The real question for teams building on agents today is no longer "is open-weight good enough?" It is "where does the routing line sit, and what does it cost to draw it in the wrong place?"
What GPT-5.6 actually changed
OpenAI unveiled GPT-5.6 on July 9, 2026 in three variants: Sol (the workhorse), Terra (mid-range), and Luna (budget).
Sol is priced at $5 input / $30 output per million tokens; Terra at $2.50/$15; and Luna at $1/$6.
CEO Sam Altman told CNBC that Sol is 54% more token-efficient on AI coding tasks than previous versions.
On the Coding Agent Index, OpenAI claims Sol "sets a new state of the art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less."
Those are vendor-reported numbers. The independent picture is narrower. On Artificial Analysis, GPT-5.6 Sol leads the Intelligence Index at 59 versus DeepSeek V4 Pro's 44, and Sol is the only one of the two with a charted coding score, ranking first on the AA Coding Agent Index at 80. A real lead - but one that costs you dearly to access.
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.
Luna deserves a closer look, because it is strategically interesting. 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. But Luna is the smallest tier with no published SWE-bench Pro score. You are not buying Sol-class capability at Luna prices; you are buying a more affordable OpenAI-ecosystem wrapper.
The open-weight field as of mid-July
Z.ai released GLM-5.2 on June 13, 2026 - a 744B-parameter Mixture-of-Experts model with a stable 1M-token context window, MIT-licensed weights, and a pitch aimed squarely at long-horizon coding.
It is the first open-weight model to top the Artificial Analysis Intelligence Index v4.1 at 51, placing fifth overall against the closed frontier.
Where DeepSeek broke through on price, GLM-5.2 is breaking through on planning quality and long-horizon coding, leading all open models on the AA Intelligence Index at 51, ahead of Nemotron 3 Ultra (48), MiniMax M3 (44), DeepSeek V4 Pro (44), and Kimi K2.6 (43) and just ~5 points below Claude Fable 5.
DeepSeek V4 occupies a different corner of the same trade-off. V4 Pro is a 1.6-trillion-parameter MoE with 49B active per forward pass; V4 Flash is a leaner 284B with 13B active, built for high-volume, cost-sensitive work.
Flash lands at 79.0% on SWE-bench Verified - within ~1.6 points of Pro's 80.6%
- and lists at $0.14/$0.28 per million tokens. DeepSeek's first-party API does retain data for training purposes , which is a real compliance constraint for some teams.
| Model | Architecture | SWE-bench | Output $/1M | License | Self-host viable? |
|---|---|---|---|---|---|
| GPT-5.6 Sol | Closed | 64.6 (Pro) | $30.00 | Proprietary | No |
| GPT-5.6 Luna | Closed | n/a | $6.00 | Proprietary | No |
| GLM-5.2 | 744B MoE, 40B active | 62.1 (Pro) | $4.40 (Z.ai) | MIT | Requires 8×H200 |
| DeepSeek V4 Pro | 1.6T MoE, 49B active | 80.6 (Verified)* | $0.87 | MIT | Requires similar |
| DeepSeek V4 Flash | 284B MoE, 13B active | 79.0 (Verified)* | $0.28 | MIT | Requires similar |
*Self-reported by DeepSeek; not yet independently verified on the same harness as Sol.
Sol is too new to be charted on the independent SWE-bench Verified leaderboard as of mid-July 2026, and DeepSeek's widely quoted 80.6% is a self-reported figure run on DeepSeek's own harness. So there is no clean independent head-to-head on that benchmark. Read the table as directional, not definitive.
The self-hosting trap
Kimi K2.5 and GLM-5.2 require multi-GPU setups with A100 or H100-class GPUs. The hardware numbers for GLM-5.2 are worth stating plainly. To serve GLM-5.2 at full quality you need roughly 750 GB of GPU memory for the official FP8 weights, which in practice means a single node of 8× NVIDIA H200 SXM5 (1,128 GB total) once you leave headroom for KV cache and activations.
Only ~40B parameters activate per forward pass, but all 744B must live in GPU memory - expert routing cannot page weights in and out at inference time without prohibitive latency.
The cloud bill is not abstract. At approximately $41,000-$42,000 per month for a production-grade AWS or GCP deployment, self-hosting GLM-5.2 is not cheap infrastructure.
Using the Z.ai hosted API at a typical 60/40 input-output mix works out to a blended price of $2.60 per million tokens , which means self-hosting breaks even only at genuinely industrial scale. One scenario that flips this: data sovereignty. If your code legally cannot leave your machines - and the GLM-5.2 API is hosted in China, which several outlets flagged as a compliance concern - then local is the only option, and the MIT license is exactly why GLM-5.2 is the model you would self-host.
For most teams, the answer is neither self-host nor go all-in on Sol. It is a routing layer.
Beagle in action#engineering, 11:02amHow to route in practice
The clearest frame is task complexity, not model tier. The serious AI stack is becoming a portfolio: frontier closed models for the hardest tasks, cheaper hosted open-weight APIs for high-volume work, local models for privacy and fallback. For draft patches, test generation, repository search, and low-risk refactors, open-weight models can already be strong alternatives. For the hardest architecture decisions, ambiguous debugging, security-sensitive changes, and final review, closed frontier models may still be worth the premium.
A routing strategy that works in practice:
Classification step first. A lightweight model (Qwen3-Coder-Next at 3B active, or even a small encoder) scores incoming tasks on a 1-3 complexity scale. A classification step at the start of each workflow decides which model handles the request, cutting costs by 60-80% while maintaining quality where it matters.
V4 Flash for the bulk. DeepSeek V4 Flash is the first open-weight model that teams immediately dropped into real agentic pipelines as a plausible substitute for an Anthropic- or OpenAI-class frontier model. The larger V4 Pro variant set the ceiling with a score of 80.6% on SWE-bench Verified, but it is Flash that broke through, capturing most of that capability at a price on the Pareto frontier of performance and cost.
GLM-5.2 via API for hard coding tasks. High reasoning mode is the right default for interactive coding agents; Max mode makes sense for batch agentic pipelines where latency is less critical than quality on hard problems.
Sol or Fable for the 10% that needs it. Ambiguous multi-system debugging, security review, and high-stakes customer-facing generation. The 30× output premium is defensible on a small fraction of calls.
Build your own evals. The only ground truth is testing models against your own tasks. Leaderboard numbers are directional; your workload is specific.
A teammate like Beagle, handling routine Slack lookups and internal Q&A at the edge of a larger agent stack, is exactly the kind of workload that fits the cheap tier - leaving Sol for the calls that genuinely need it.
Open-weight models vs GPT-5.6: common questions
Is GPT-5.6 Sol worth the premium for coding agents?
Sol leads independent benchmarks and carries real advantages in multimodal input and OpenAI ecosystem tooling. For the hardest 10% of engineering tasks - ambiguous debugging, security-critical generation, multi-system reasoning - the capability premium is defensible. For high-volume, routine coding work, the 30× output cost difference points toward GLM-5.2 or DeepSeek V4 Flash.
How does GLM-5.2 compare to GPT-5.6 on real tasks?
On SWE-bench Pro, GPT-5.6 Sol scores 64.6 versus GLM-5.2's 62.1 - a 2.5-point gap. On the Artificial Analysis Intelligence Index, Sol leads at 59 versus GLM-5.2's 51. GLM-5.2 is MIT-licensed, serves at roughly one-sixth the cost of Sol, and leads open weights on Artificial Analysis' real-world agentic benchmark, effectively level with GPT-5.5 xhigh. The gap is real; it is not large.
Can my team self-host GLM-5.2 to save money?
The break-even token volume for self-hosting GLM-5.2 on cloud GPUs is roughly 15 billion tokens per month. For everyone below that, the Z.ai hosted API at $1.40/$4.40 per million tokens provides near-identical performance at a fraction of the operational complexity. The only strong reason to self-host below that threshold is data sovereignty - a real need, not a general recommendation.
Should teams use DeepSeek V4 if data privacy matters?
Pick DeepSeek V4 when open weights, self-hosting or weight inspection, and extremely low token prices matter more than closed-service frontier features. If your compliance policy prohibits data leaving your infrastructure, running DeepSeek weights on your own cloud account solves the residency problem. Note that DeepSeek's first-party API retains data for training purposes
- self-hosting the MIT weights removes that concern entirely.
What does GPT-5.6 Luna actually buy you?
Luna ($1 input / $6 output) is OpenAI's first model priced inside the open-weight band. Luna and GLM-5.2 cost roughly the same on a blended workload. 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 Sol. Luna's value is ecosystem access - ChatGPT Work, Codex, native OpenAI tool-calling - not raw coding performance.