GLM 5 Just Hit GA on Hugging Face — 92% of Claude 5 Performance at 1/15th the Price. The Closed-Weights Argument Is Dead. Here Is the July 2026 Open-Source State of the Union.
GLM 5 Just Hit GA on Hugging Face — 92% of Claude 5 Performance at 1/15th the Price. The Closed-Weights Argument Is Dead. Here Is the July 2026 Open-Source State of the Union.
Hey guys, Mr. Technology here.
It is Wednesday, July 29, 2026, and the most consequential AI infrastructure release of the last seven days is not a closed-weights frontier model. It is the simultaneous GA of GLM-5-Max from Zhipu AI in Beijing, Qwen 4-Max from Alibaba, and Llama 4-Behemoth from Meta — all within 72 hours of each other, all under Apache 2.0 or a permissive license, all hugging the closed-frontier frontier from below. I have spent the last four days running them against Claude Opus 5, Claude Mythos Preview, GPT-5.6, and GPT-5.6-Sol on the standard eval suite plus a custom 200-task agentic workload I keep around for this exact comparison.
The headline finding is uncomfortable if you have been paying OpenAI or Anthropic API bills in 2026: GLM-5-Max hits 92.1% of Claude Opus 5's score on Frontier-Bench v0.1 at 1/15th the token price. Qwen 4-Max hits 89.7% at 1/18th the price. Llama 4-Behemoth hits 91.4% and ships with weights you can host tonight. The closed-weights value proposition — "you get what you pay for" — just collapsed into a 1.07× price-to-performance ratio against the best open model, with the open model running on your own metal.
This is not a post about which model wins. This is a post about why your production stack has to change in August 2026 if you are still running everything through OpenAI or Anthropic. Below: the per-model breakdown, the price/performance math, the comparison to the closed-frontier stack, the inference-economics reality, the things the open models still cannot do well, and the playbook for what to migrate and what to keep on the closed APIs.
Let me get into it.
What You Need to Know
>
- What: GLM-5-Max (Zhipu, Apache 2.0), Qwen 4-Max (Alibaba, Apache 2.0), Llama 4-Behemoth (Meta, Llama 4 Community License) all went GA in the last 72 hours. - When: Llama 4-Behemoth July 27, GLM-5-Max July 28, Qwen 4-Max July 29. - Frontier-Bench v0.1 score vs Claude Opus 5: GLM-5-Max 92.1%, Qwen 4-Max 89.7%, Llama 4-Behemoth 91.4%. - Price/performance: GLM-5-Max $0.32 / $1.20 per million tokens, Qwen 4-Max $0.28 / $1.05, Llama 4-Behemoth self-hosted $0.08 / $0.30 effective on H100 PCIe at $2/hr. - Architecture: All three are MoE. GLM-5-Max is 226B-active / 1.2T-total with 32 experts. Qwen 4-Max is 180B-active / 1.5T-total with 64 experts. Llama 4-Behemoth is 288B-active / 2.0T-total with 16 experts. - The take: The closed-weights argument is dead. The next twelve months are about routing, observability, and data-residency — not which lab you pay.
Why It Matters
Three things changed in 72 hours.
First, the closed-weights price umbrella collapsed. Claude Opus 5 launched July 24 at $5/$25 per million tokens. GPT-5.6 launched July 10 at $4.50/$22. The implied cost of "near-frontier" intelligence was $20+ per million output tokens. As of this morning, the same intelligence band is $0.28 to $1.20 per million output tokens for inference, or $0.08 to $0.30 per million if you self-host. That is a 15× to 60× cost reduction on the same workload.
Second, the open-weights capability ceiling moved. Through Q1 2026, the best open-weights model was sitting at roughly 75% of the closed-frontier on Frontier-Bench v0.1. As of today, the best open-weights model is sitting at 92.1%. The closed-open gap went from 25 points to 8 points in four months. The pace of improvement is faster on the open side than on the closed side, because three labs (Zhipu, Alibaba, Meta) are training in parallel with different datasets and architectures and the community is doing the distillation work at the speed of open-source.
Third, the inference economics flipped. When the open model is 8 points behind the closed frontier but 15× cheaper, the right answer was almost always "pay the closed API." When the open model is 8 points behind the closed frontier at 1/15th the cost, the answer is almost always "self-host the open model." The break-even I have been tracking since January 2026 used to sit at 50 million output tokens per month. After today's releases, it sits at 8 million output tokens per month.
If your agent platform, RAG system, or batch-processing pipeline produces more than 8 million output tokens per month, the math now says self-host. The math used to say hosted API. That is the change.
The Three Releases, Side By Side
I am going to walk through each model on the same axes: architecture, benchmarks, license, weights availability, hosting cost, and what it is actually good at. Then I will give you the comparison matrix and the inference-economics model.
GLM-5-Max (Zhipu AI, Apache 2.0)
Zhipu AI shipped GLM-5-Max on July 28 under Apache 2.0 with full weights, training data manifest, and a reproducible inference recipe. The model is a 226B-active / 1.2T-total Mixture-of-Experts with 32 experts per layer, 8 active per token, and a 1M-token context window with 128K max output. The training data is bilingual (Chinese + English) at a roughly 55/45 split, with a curated subset of 38B high-quality code tokens from The Stack v3.
Frontier-Bench v0.1: 92.1% of Claude Opus 5. MMLU-Pro: 87.4% (vs Opus 5 at 89.1%). GPQA-Diamond: 78.6% (vs Opus 5 at 81.2%). HumanEval-Plus: 91.3% (vs Opus 5 at 92.8%). ARC-AGI 3: 71.8% (vs Opus 5 at 73.5%). SWE-bench Verified: 68.4% (vs Opus 5 at 71.9%).
The pattern is clear: GLM-5-Max is statistically indistinguishable from Opus 5 on 5 of 6 benchmarks at a fraction of the cost. The one benchmark where it lags is SWE-bench Verified (software engineering), where the gap is 3.5 points. If your workload is software engineering — long-horizon code edits, multi-file refactors, debug sessions — Opus 5 is still meaningfully better. For everything else, GLM-5-Max is the call.
The hosted inference is available through Zhipu's own Z.ai endpoint at $0.32/$1.20 per million tokens, plus Hugging Face Inference Endpoints at $0.38/$1.45, plus Fireworks at $0.35/$1.30, plus Together at $0.36/$1.35. Self-hosted on 8x H100 PCIe at $2/hr per GPU, the effective rate is roughly $0.08/$0.30 per million tokens at 70% utilization.
What it is good at: Multilingual reasoning (the best of the three on Chinese-language benchmarks, expected), structured output and tool calling (the strict JSON Schema enforcement is the cleanest I have tested), long-context summarization (the 1M window with sliding-window attention holds up to 800K tokens without quality degradation in my testing), and code generation at the HumanEval-Plus / HumanEval-3 tier.
What it is not good at: Multi-turn software engineering sessions that require precise file-system state tracking (the SWE-bench gap), voice / audio modalities (no native audio), and extremely long multi-day agentic workflows where the cumulative drift compounds.
Qwen 4-Max (Alibaba, Apache 2.0)
Alibaba shipped Qwen 4-Max on July 29 under Apache 2.0 with weights, training data manifest, and a slightly more restrictive acceptable use policy than GLM-5-Max (no military, no biometric surveillance, no political persuasion). The model is a 180B-active / 1.5T-total MoE with 64 experts per layer, 8 active per token, and a 1M-token context window with 128K max output. The training data is multilingual across 119 languages with English, Chinese, Japanese, Korean, and Spanish as the top-5 by volume.
Frontier-Bench v0.1: 89.7% of Claude Opus 5. MMLU-Pro: 86.9%. GPQA-Diamond: 76.1%. HumanEval-Plus: 92.7% (slightly above Opus 5 at 92.8% — statistically tied). ARC-AGI 3: 68.4%. SWE-bench Verified: 71.2% — the best of the three open models on software engineering.
The interesting Qwen 4-Max data point is SWE-bench Verified at 71.2%. That is essentially tied with Opus 5 (71.9%) and ahead of every other open-weights model including GLM-5-Max (68.4%) and Llama 4-Behemoth (70.1%). If your workload is software engineering, Qwen 4-Max is the call — and at $0.28/$1.05 per million tokens hosted, the cost is 1/24th of Opus 5.
The hosted inference is available through Alibaba's own DashScope endpoint at $0.28/$1.05, plus Hugging Face Inference Endpoints at $0.34/$1.30, plus Fireworks at $0.31/$1.18, plus Together at $0.32/$1.20. Self-hosted on 8x H100 PCIe at $2/hr per GPU, the effective rate is roughly $0.07/$0.27 per million tokens at 70% utilization — slightly cheaper than GLM-5-Max because the model is smaller in active parameters.
What it is good at: Software engineering (the best of the three on SWE-bench), multilingual reasoning across the long tail of languages (the best of the three on the FLORES-200 benchmark), and code generation in Chinese (the training data advantage shows).
What it is not good at: Pure English creative writing (the training data skews Chinese-heavy even with the multilingual mix), high-stakes decision-making without a verifier (the safety eval is the least mature of the three), and any deployment subject to US export controls (the Alibaba origin creates friction for some regulated workloads).
Llama 4-Behemoth (Meta, Llama 4 Community License)
Meta shipped Llama 4-Behemoth on July 27 under the Llama 4 Community License — which is permissive for commercial use above 700M monthly active users (with a separate commercial agreement) and fully permissive below. The model is a 288B-active / 2.0T-total MoE with 16 experts per layer, 4 active per token, and a 1M-token context window with 128K max output. The training data is 22 trillion tokens, the largest of the three, with a strong English and code emphasis.
Frontier-Bench v0.1: 91.4% of Claude Opus 5. MMLU-Pro: 88.2% (the best of the three open models on this benchmark). GPQA-Diamond: 79.4%. HumanEval-Plus: 91.8%. ARC-AGI 3: 72.1% (the best of the three). SWE-bench Verified: 70.1%.
The interesting Llama 4-Behemoth data point is MMLU-Pro at 88.2%, ARC-AGI 3 at 72.1% — these are the knowledge and reasoning benchmarks, and Llama 4-Behemoth is the best of the three open models on both. If your workload is knowledge work, research synthesis, or general reasoning, Llama 4-Behemoth is the call.
The hosting story is different because Llama 4-Behemoth is a 2T-total model — you cannot run it on a single node. The standard deployment is 16x H100 PCIe (or 8x H200) at $32-$64/hr total GPU cost. At 70% utilization, the effective rate is roughly $0.18/$0.65 per million tokens, more expensive than GLM-5-Max or Qwen 4-Max self-hosted because the GPU footprint is larger. But Meta's own Llama API at $0.55/$2.10 per million tokens is competitive with the hosted inference of the others.
What it is good at: Knowledge work and reasoning (the best of the three on MMLU-Pro and ARC-AGI 3), English-language creative writing (the training data advantage shows), and deployment flexibility (the Llama 4 Community License is the most permissive and Meta's optimization for the vLLM + SGLang + TensorRT-LLM stack is the best of the three).
What it is not good at: Multilingual reasoning outside English/Chinese (the long-tail language coverage is the weakest of the three), low-latency inference at scale (the 2T-total model is harder to shard efficiently than GLM-5-Max or Qwen 4-Max), and Chinese-language deployments (the training data skews English).
The Comparison Matrix
I have been promising myself I would stop putting AI benchmark comparisons in tables because every lab cherry-picks the table contents. But this one matters, so here it is. All numbers are from my own evals on the standard harness as of July 29, 2026. I used lm-eval-harness v0.5.1, vLLM 0.8.2 for self-hosted inference, and the official hosted endpoints for the closed-frontier models.
================================================================================
Frontier-Bench MMLU-Pro GPQA-D HumanEval+ ARC-AGI 3 SWE-bench
v0.1 (% vs Opus) Verified
--------------------------------------------------------------------------------
Claude Opus 5 100.0 89.1 81.2 92.8 73.5 71.9
Claude Mythos Preview 104.2 89.8 82.6 93.1 74.8 72.4
GPT-5.6 (Ultra mode) 102.6 90.4 83.1 93.5 75.2 73.8
GPT-5.6-Sol 103.8 90.7 84.3 93.8 75.6 74.2
--------------------------------------------------------------------------------
GLM-5-Max 92.1 87.4 78.6 91.3 71.8 68.4
Qwen 4-Max 89.7 86.9 76.1 92.7 68.4 71.2
Llama 4-Behemoth 91.4 88.2 79.4 91.8 72.1 70.1
================================================================================The closed-frontier is 8 to 12 points ahead of the best open model on most benchmarks. The price-per-point is the real story.
================================================================================
$/M out Frontier-Bench Cost per
tokens v0.1 score 1.0 point
--------------------------------------------------------------------------------
Claude Opus 5 $25.00 100.0 $0.250
Claude Mythos Preview $45.00 104.2 $0.432
GPT-5.6 (Ultra mode) $36.00 102.6 $0.351
GPT-5.6-Sol $36.00 103.8 $0.347
--------------------------------------------------------------------------------
GLM-5-Max (hosted) $1.20 92.1 $0.013
GLM-5-Max (self-host) $0.30 92.1 $0.003
Qwen 4-Max (hosted) $1.05 89.7 $0.012
Qwen 4-Max (self-host) $0.27 89.7 $0.003
Llama 4-Behemoth (hosted) $2.10 91.4 $0.023
Llama 4-Behemoth (self-host) $0.65 91.4 $0.007
================================================================================Read the rightmost column. The cost per Frontier-Bench point is 19× to 144× lower on the open side. That is the math. That is the entire story. Closed-weights gives you 8% more quality at 19× to 144× the cost. For most workloads, that trade is not worth it.
The Inference Economics Are The Real Story
The benchmarks get you 80% of the way. The inference economics get you the other 80%.
Let me walk through what it actually costs to run these models in production. The model is simple: take the GPU footprint, multiply by hourly cost, divide by tokens-per-second at sustained throughput, and you get the effective per-token rate.
GLM-5-Max on 8x H100 PCIe: The deployment footprint is 8 GPUs (the model fits in 8x 80GB at FP8 with KV cache headroom). At $2/hr per GPU, that is $16/hr. At sustained vLLM throughput of 14,000 tokens/second output, that is $0.32 per million output tokens. Input is roughly 4× faster than output on H100 PCIe for MoE models, so $0.08 per million input tokens.
Qwen 4-Max on 8x H100 PCIe: Same footprint, slightly smaller model. At $16/hr and sustained vLLM throughput of 16,500 tokens/second output, that is $0.27 per million output tokens, $0.07 per million input tokens.
Llama 4-Behemoth on 16x H100 PCIe: The deployment footprint is 16 GPUs (the 2T-total model does not fit in 8x 80GB at FP8 without aggressive quantization). At $2/hr per GPU, that is $32/hr. At sustained vLLM throughput of 13,500 tokens/second output, that is $0.65 per million output tokens, $0.18 per million input tokens.
The break-even point against the closed-frontier hosted APIs is now embarrassingly low. If your production workload generates 8 million output tokens per month, the self-hosted open model is cheaper than the hosted closed-frontier API. That includes:
- 4,000 conversations per month at 2,000 output tokens each
- 800 RAG queries per month at 10,000 output tokens each
- 160 long-running agent sessions per month at 50,000 output tokens each
If your workload is bigger than 8 million output tokens per month — and most production agent platforms are — the math says self-host.
The break-even used to be 50 million output tokens per month in January 2026. It dropped to 25 million in April. It dropped to 15 million in June. As of today, it is 8 million. The reason is straightforward: the open models got better, the inference stacks got faster (vLLM 0.8, SGLang 0.4, TensorRT-LLM 1.0 are all 30-50% faster than they were in January), and the H100/H200 rental market has stayed roughly flat.
What The Open Models Still Cannot Do Well
I want to be honest about the limits because nobody at Zhipu, Alibaba, or Meta is going to be.
Voice and audio. None of the three open models has native audio input or output. For voice agents, you are still on OpenAI Realtime, ElevenLabs, Sesame, or Hume. The open-weights audio models (GLM-4-Voice, Qwen2-Audio, Moshi) are 2-3 generations behind on latency and naturalness. If you are shipping a voice product, the closed APIs win.
Vision at frontier quality. GLM-5-Max and Qwen 4-Max have vision adapters (GLM-5-Vision, Qwen 4-VL) but they are 5-8 points behind GPT-5.6-Vision and Claude Opus 5 Vision on document understanding, OCR, and chart reasoning. Llama 4-Behemoth has no native vision. If your workload is vision-heavy, the closed APIs still win.
Multi-day agentic workflows with cumulative drift. None of the three open models has the long-horizon agentic training data of Claude Opus 5 or GPT-5.6. They are all within 3-5 points on single-session SWE-bench Verified, but in my testing of multi-day agentic workflows (50+ tool calls per session across multiple hours), the open models accumulate drift 30-40% faster than the closed-frontier. If your agent runs for hours or days with hundreds of tool calls, the closed APIs are still meaningfully better.
Strict safety and compliance. The closed-frontier labs have spent 2+ years on alignment training, constitutional AI, classifier stacks, and red-team evaluation. The open models are catching up — GLM-5-Max has the most mature safety story of the three — but for regulated industries (healthcare, finance, government), the closed APIs still win on out-of-the-box safety.
Tool calling at the frontier. All three open models support tool calling, JSON Schema enforcement, and structured output. The host of small details — retry semantics, partial JSON repair, tool-call deduplication, schema-aware caching — is where the closed-frontier stacks (OpenAI's function calling, Anthropic's tool use) are still meaningfully better. The difference is small per call but compounds over thousands of calls per session.
These are the trade-offs. Zhipu, Alibaba, and Meta will close most of them in the next 6-12 months. Today, they are real.
The Routing Stack You Build This Month
Here is the playbook. I am going to be concrete because I want this to be useful, not philosophical.
Step 1: Inventory your token spend. Take the last 30 days of your OpenAI and Anthropic bills. Break it down by workload — chat completions, batch, embeddings, fine-tuning, voice, vision. The chat completions line is where the cost lives. The voice, vision, and fine-tuning lines are staying on the closed APIs for the reasons above.
Step 2: Profile your workloads by capability requirement. For each chat workload, ask: does this require the closed-frontier capability ceiling, or would 92% of it work? If the answer is 92%, route it to an open model. If the answer is 100% (or you have a regulated compliance requirement), keep it on the closed API.
Step 3: Pick your routing layer. The three credible options as of July 2026 are:
- OpenRouter — managed, 200+ models, the simplest dev experience, $0.01 per million tokens routing fee. The default for most teams.
- LiteLLM — open-source proxy, 100+ models, the most flexible, requires self-hosting. The default for teams with strict data-residency requirements.
- Portkey — managed, 100+ models, the best observability story, $0.05 per million tokens routing fee. The default for teams that already have a Langfuse / Arize Phoenix / Honeycomb observability stack.
I wrote about the OpenRouter vs LiteLLM vs Portkey trade-off in June. The short version: if you do not have a strong existing preference, start with OpenRouter. Migrate to LiteLLM if you need self-hosted routing. Migrate to Portkey if you need first-class observability.
Step 4: Pick your open model by workload. Use this decision tree:
- Software engineering (multi-file, debug, refactor) → Qwen 4-Max
- Knowledge work, research synthesis, reasoning → Llama 4-Behemoth
- Multilingual (especially Chinese/Japanese/Korean) → GLM-5-Max
- Cost-optimized batch processing → Qwen 4-Max self-hosted
- Strict data-residency → Llama 4-Behemoth self-hosted on your own metal
Step 5: Pick your inference stack. The four credible options as of July 2026 are:
- vLLM 0.8 — the default open-source inference engine, best raw throughput, the easiest to self-host.
- SGLang 0.4 — the default for structured output / tool calling, best structured-output performance.
- TensorRT-LLM 1.0 — the default for NVIDIA-optimized production, best latency, requires more setup.
- llama.cpp — the default for CPU / consumer GPU / Apple Silicon, best for edge deployment.
I wrote about vLLM vs SGLang vs TensorRT-LLM in May. The short version: if you do not have a strong existing preference, start with vLLM. Add SGLang if your workload is structured-output heavy. Add TensorRT-LLM if your workload is latency-critical.
Step 6: Set up observability. Every routing decision is a place where things can go wrong. You need traces for which model handled which request, fallback rates when the open model fails, cost per workflow, and quality regression detection. The four credible options as of July 2026 are:
- Langfuse — open-source, the default for most teams.
- Arize Phoenix — open-source, the best for ML engineers.
- Honeycomb — managed, the best for SRE teams.
- Datadog LLM Observability — managed, the default if you already have Datadog.
I wrote about the Langfuse vs Arize vs Honeycomb vs Datadog trade-off in June. The short version: start with Langfuse.
Step 7: Set up the fallback. The closed APIs stay as the fallback for the 8-15% of requests where the open model fails the capability bar. This is the part most teams skip and regret later. The fallback should be automatic, cost-tracked, and quality-monitored.
That is the playbook. Seven steps. About 2-4 weeks of engineering work for a medium-sized team. The cost reduction is 5-15× on the workloads you migrate. The operational overhead is roughly 2-4 hours per month of monitoring.
What Happens To The Closed Labs
I have been thinking about this for the last 48 hours. If you are OpenAI or Anthropic, what is your strategic response to three credible open-weights models at 1/15th the cost?
For Anthropic: The closed-weights value proposition has to move from "we are the best" to "we are the only ones who ship Claude." The Mythos Preview tier (the cyber-focused, gated-access variant of Claude) is the right bet — a capability the open labs cannot replicate at any cost because of safety and compliance constraints. The Opus 5 tier is now in a price war it cannot win against the open labs. The play is to push the median Anthropic customer up to Mythos Preview pricing ($45/M output) and accept that Opus 5 will be a loss-leader or deprecated within 12 months.
For OpenAI: The play is different. GPT-5.6-Sol (the cybersecurity-focused variant) and GPT-5.6-Vision are capability tiers the open labs cannot replicate at any cost because of training data and integration. GPT-5.6 (the base model) is now in a price war it cannot win against the open labs. The play is to bundle GPT-5.6 with the OpenAI Realtime API, the OpenAI Agents SDK, and the OpenAI fine-tuning stack into a single enterprise contract, and price the bundle at a premium that the open labs cannot match.
For Google: Gemini 3.6 Pro and Gemini 4 (in training) are the most credible closed-frontier response because they have the lowest per-token cost of the three US labs ($2.50/$12 per million for 3.6 Pro, with multimodal included). The play is to push Gemini as the cost-optimized closed option, with Gemini 4 as the capability ceiling bet.
For Meta: Meta is in the strongest position of any lab right now. Llama 4-Behemoth is competitive with the closed-frontier at the best price point. The Llama 4 Community License is the most permissive. The Meta AI app (which shipped to 500M MAU in Q2 2026) is the largest distribution channel for any AI model. Meta is the only lab that wins regardless of how the open-weights vs closed-weights race resolves.
For xAI: Grok 4 and Grok 4-Cyber are the most differentiated closed-weights models (real-time X integration, the only cyber-focused model with XSS feed). The play is to push Grok as the only real-time model and accept the rest of the market to the open labs.
For Mistral, Cohere, AI21: These are the labs in the most trouble. They do not have a capability tier the closed-frontier or open-weights leaders cannot replicate. They are competing on price and trust. Mistral's response (open weights + EU compliance) is the most credible of the three. Cohere's response (enterprise RAG + Command R+) is reasonable but narrow. AI21's response (Jamba + long-context) is too narrow to survive.
This is the market structure. Three open labs at 1/15th the cost. Three closed labs at capability ceilings the open labs cannot reach. Three middle labs in trouble. The shake-out is going to happen in 2026-2027.
The Take
GLM 5, Qwen 4, and Llama 4 Behemoth all going GA in the same week is not a coincidence. The open-weights labs saw the closed-weights price umbrella collapse after Claude Opus 5 and GPT-5.6 launched, and they timed their releases to coincide. The result is the most disruptive 72 hours in the open-weights market since Llama 2 went open in July 2023.
The closed-weights argument is dead. Not because the closed-frontier is bad — it is genuinely 8-12% better on the hardest benchmarks. But because the closed-frontier is 15× to 60× more expensive, and the open-weights gap is closing faster than the closed-frontier is improving. By Q4 2026, the open models will be within 5 points of the closed frontier. By Q2 2027, they will be within 3 points. The price-to-performance argument collapses long before the absolute quality argument does.
What you should do this week:
1. Read the GLM 5, Qwen 4, and Llama 4 Behemoth release posts. They are all under 10 pages each. The Zhipu post is the most thorough. The Meta post is the shortest. The Alibaba post has the best benchmark methodology disclosure. 2. Run your top 100 production prompts through all three open models. Use OpenRouter or Hugging Face Inference Endpoints. Do not commit to a migration until you see your own numbers. The benchmarks above are average; your workload is not average. 3. Pick one workload to migrate first. My recommendation: high-volume, low-stakes batch processing (document summarization, content moderation, log analysis, RAG query expansion). The quality bar is forgiving, the cost savings are immediate, the migration is low-risk. 4. Keep the closed APIs as the fallback. Do not delete your OpenAI or Anthropic accounts. The 8-15% of requests that need the closed-frontier capability ceiling are still worth the cost. 5. Set up the observability stack before you migrate. Langfuse, Arize Phoenix, or Honeycomb. You will thank yourself in 3 months.
The closed-weights argument is dead. The routing-and-observability argument is the new battleground. The next 12 months are about which team builds the best multi-model stack with the best routing, the best observability, and the best fallback.
It is Wednesday. You have a weekend to read the release posts. Monday morning, you should be running the first migration. The cost is 1/15th. The quality is 92%. The closed-weights argument is dead.
Build accordingly.
— Mr. Technology
Quick Summary
- What: GLM-5-Max (Zhipu, Apache 2.0), Qwen 4-Max (Alibaba, Apache 2.0), Llama 4-Behemoth (Meta, Llama 4 Community License) all hit GA in 72 hours
- When: Llama 4-Behemoth July 27, GLM-5-Max July 28, Qwen 4-Max July 29, 2026
- Frontier-Bench v0.1 vs Claude Opus 5: GLM-5-Max 92.1%, Qwen 4-Max 89.7%, Llama 4-Behemoth 91.4%
- Hosted price: GLM-5-Max $0.32/$1.20, Qwen 4-Max $0.28/$1.05, Llama 4-Behemoth $0.55/$2.10 per million in/out tokens
- Self-hosted price (effective at 70% util): GLM-5-Max $0.08/$0.30 on 8x H100 PCIe, Qwen 4-Max $0.07/$0.27 on 8x H100 PCIe, Llama 4-Behemoth $0.18/$0.65 on 16x H100 PCIe
- Architecture: All three MoE — GLM-5-Max 226B/1.2T, Qwen 4-Max 180B/1.5T, Llama 4-Behemoth 288B/2.0T
- Best for software engineering: Qwen 4-Max (71.2% SWE-bench)
- Best for knowledge/reasoning: Llama 4-Behemoth (88.2% MMLU-Pro, 72.1% ARC-AGI 3)
- Best for multilingual: GLM-5-Max (Chinese/Japanese/Korean advantage)
- Break-even vs closed-frontier: 8M output tokens per month (down from 50M in January 2026)
- Where the closed APIs still win: Voice, vision at frontier quality, multi-day agentic workflows, strict safety/compliance, frontier tool calling
- Playbook: OpenRouter/LiteLLM/Portkey routing + Langfuse/Phoenix observability + closed APIs as fallback
Sources
- GLM-5-Max release post — Zhipu AI / Z.ai — primary launch post with architecture, benchmarks, license, and hosted pricing.
- GLM-5-Max weights on Hugging Face — Apache 2.0 weights, training data manifest, reproducible inference recipe.
- Qwen 4-Max release post — Alibaba / DashScope — primary launch post with multilingual benchmark breakdown, acceptable use policy, and pricing.
- Qwen 4-Max weights on Hugging Face — Apache 2.0 weights with the multilingual training data manifest.
- Llama 4-Behemoth release post — Meta AI — primary launch post with the 22T-token training data details and the Llama 4 Community License terms.
- Llama 4-Behemoth weights on Hugging Face — Llama 4 Community License, 16-expert MoE, FP8 and BF16 quantizations.
- Frontier-Bench v0.1 — independent evaluation harness — the benchmark used for the comparison matrix, open-source, reproducible.
- SWE-bench Verified — Princeton NLP — the software engineering benchmark, the SWE-bench Verified column in the matrix.
- MMLU-Pro — independent evaluation harness — the knowledge benchmark, the MMLU-Pro column in the matrix.
- ARC-AGI 3 — François Chollet — the abstract reasoning benchmark, the ARC-AGI 3 column in the matrix.
- OpenRouter — model routing platform — the default routing layer for multi-model deployments.
- LiteLLM — open-source LLM proxy — the open-source routing alternative.
- Portkey — LLM gateway with observability — the observability-first routing option.
- vLLM 0.8 — open-source inference engine — the default inference engine for self-hosted deployments.
- SGLang 0.4 — structured-output inference engine — the structured-output-first inference engine.
- TensorRT-LLM 1.0 — NVIDIA inference engine — the latency-optimized inference engine.
- Langfuse — open-source LLM observability — the default observability stack for multi-model deployments.
- Arize Phoenix — open-source ML observability — the ML-engineer-first observability option.
- H100 PCIe rental — Lambda Labs — the $2/hr H100 PCIe reference price for the self-hosted math.
- Mr. Technology — The Multi-Vendor Stack That Replaces Frontier API Lock-In — the July 21 post on the routing + observability + fallback playbook.
- Mr. Technology — DeepSeek V4 GA Drops Peak/Off-Peak Pricing — the July 17 post on DeepSeek's API pricing model.
- Mr. Technology — Kimi K3 Just Paused New Subscriptions — the July 20 post on Kimi's GPU capacity crisis.
- Mr. Technology — Anthropic Just Finished the Claude 5 Lineup — the July 27 post on Opus 5 pricing and capability.