- Open source AI accelerates as GLM-5.1 brings capable coding to everyone.
Is China picking back up the open source AI baton?
Z.ai, also known as Zhupai AI, a Chinese AI startup best known for its powerful, open source GLM family of models, has unveiled GLM-5.1 today under a permissive MIT License, allowing for enterprises to download, customize and use it for commercial purposes. They can do so on Hugging Face.
This follows its release of GLM-5 Turbo, a faster version, under only proprietary license last month.
The new GLM-5.1 is designed to work autonomously for up to eight hours on a single task, marking a definitive shift from vibe coding to agentic engineering.
The release represents a pivotal moment in the evolution of artificial intelligence. While competitors have focused on increasing reasoning tokens for better logic, Z.ai is optimizing for productive horizons.
GLM-5.1 is a 754-billion parameter Mixture-of-Experts model engineered to maintain goal alignment over extended execution traces that span thousands of tool calls.
"agents could do about 20 steps by the end of last year," wrote z.ai leader Lou on X. "glm-5.1 can do 1,700 rn. autonomous work time may be the most important curve after scaling laws. glm-5.1 will be the first point on that curve that the open-source community can verify with their own hands. hope y'all like it^^"
In a market increasingly crowded with fast models, Z.ai is betting on the marathon runner. The company, which listed on the Hong Kong Stock Exchange in early 2026 with a market capitalization of $52.83 billion, is using this release to cement its position as the leading independent developer of large language models in the region.
Technology: the staircase pattern of optimization
GLM-5.1s core technological breakthrough isn't just its scale, though its 754 billion parameters and 202,752 token context window are formidable, but its ability to avoid the plateau effect seen in previous models.
In traditional agentic workflows, a model typically applies a few familiar techniques for quick initial gains and then stalls. Giving it more time or more tool calls usually results in diminishing returns or strategy drift.
Z.ai research demonstrates that GLM-5.1 operates via what they call a staircase pattern, characterized by periods of incremental tuning within a fixed strategy punctuated by structural changes that shift the performance frontier.
In Scenario 1 of their technical report, the model was tasked with optimizing a high-performance vector database, a challenge known as VectorDBBench.
The model is provided with a Rust skeleton and empty implementation stubs, then uses tool-call-based agents to edit code, compile, test, and profile. While previous state-of-the-art results from models like Claude Opus 4.6 reached a performance ceiling of 3,547 queries per second, GLM-5.1 ran through 655 iterations and over 6,000 tool calls. The optimization trajectory was not linear but punctuated by structural breakthroughs.
At iteration 90, the model shifted from full-corpus scanning to IVF cluster probing with f16 vector compression, which reduced per-vector bandwidth from 512 bytes to 256 bytes and jumped performance to 6,400 queries per second.
By iteration 240, it autonomously introduced a two-stage pipeline involving u8 prescoring and f16 reranking, reaching 13,400 queries per second. Ultimately, the model identified and cleared six structural bottlenecks, including hierarchical routing via super-clusters and quantized routing using centroid scoring via VNNI. These efforts culminated in a final result of 21,500 queries per second, roughly six times the best result achieved in a single 50-turn session.
This demonstrates a model that functions as its own research and development department, breaking complex problems down and running experiments with real precision.
The model also managed complex execution tightening, lowering scheduling overhead and improving cache locality. During the optimization of the Approximate Nearest Neighbor search, the model proactively removed nested parallelism in favor of a redesign using per-query single-threading and outer concurrency.
When the model encountered iterations where recall fell below the 95 percent threshold, it diagnosed the failure, adjusted its parameters, and implemented parameter compensation to recover the necessary accuracy. This level of autonomous correction is what separates GLM-5.1 from models that simply generate code without testing it in a live environment.
Kernelbench: pushing the machine learning frontier
The model's endurance was further tested in KernelBench Level 3, which requires end-to-end optimization of complete machine learning architectures like MobileNet, VGG, MiniGPT, and Mamba.
In this setting, the goal is to produce a faster GPU kernel than the reference PyTorch implementation while maintaining identical outputs. Each of the 50 problems runs in an isolated Docker container with one H100 GPU and is limited to 1,200 tool-use turns. Correctness and performance are evaluated against a PyTorch eager baseline in separate CUDA contexts.
The results highlight a significant performance gap between GLM-5.1 and its predecessors. While the original GLM-5 improved quickly but leveled off early at a 2.6x speedup, GLM-5.1 sustained its optimization efforts far longer. It eventually delivered a 3.6x geometric mean speedup across 50 problems, continuing to make useful progress well past 1,000 tool-use turns.
Although Claude Opus 4.6 remains the leader in this specific benchmark at 4.2x, GLM-5.1 has meaningfully extended the productive horizon for open-source models.
This capability is not simply about having a longer context window; it requires the model to maintain goal alignment over extended execution, reducing strategy drift, error accumulation, and ineffective trial and error. One of the key breakthroughs is the ability to form an autonomous experiment, analyze, and optimize loop, where the model can proactively run benchmarks, identify bottlenecks, adjust strategies, and continuously improve results through iterative refinement.
All solutions generated during this process were independently audited for benchmark exploitation, ensuring the optimizations did not rely on specific benchmark behaviors but worked with arbitrary new inputs while keeping computation on the default CUDA stream.
Product strategy: subscription and subsidies
GLM-5.1 is positioned as an engineering-grade tool rather than a consumer chatbot. To support this, Z.ai has integrated it into a comprehensive Coding Plan ecosystem designed to compete directly with high-end developer tools.
The product offering is divided into three subscription tiers, all of which include free Model Context Protocol tools for vision analysis, web search, web reader, and document reading.
The Lite tier at $27 USD per quarter is positioned for lightweight workloads and offers three times the usage of a comparable Claude Pro plan. The Pro tier at $81 per quarter is designed for complex workloads, offering five times the Lite plan usage and 40 to 60 percent faster execution.
The Max tier at $216 per quarter is aimed at advanced developers with high-volume needs, ensuring guaranteed performance during peak hours.
For those using the API directly or through platforms like OpenRouter or Requesty, Z.ai has priced GLM-5.1 at $1.40 per one million input tokens and $4.40 per million output tokens. There's also a cache discount available for $0.26 per million input tokens.
Model
Input
Output
Total Cost
Source
Grok 4.1 Fast
$0.20
$0.50
$0.70
xAI
MiniMax M2.7
$0.30
$1.20
$1.50
MiniMax
Gemini 3 Flash
$0.50
$3.00
$3.50
Google
Kimi-K2.5
$0.60
$3.00
$3.60
Moonshot
MiMo-V2-Pro (≤256K)
$1.00
$3.00
$4.00
Xiaomi MiMo
GLM-5
$1.00
$3.20
$4.20
Z.ai
GLM-5-Turbo
$1.20
$4.00
$5.20
Z.ai
GLM-5.1
$1.40
$4.40
$5.80
Z.ai
Claude Haiku 4.5
$1.00
$5.00
$6.00
Anthropic
Qwen3-Max
$1.20
$6.00
$7.20
Alibaba Cloud
Gemini 3 Pro
$2.00
$12.00
$14.00
Google
GPT-5.2
$1.75
$14.00
$15.75
OpenAI
GPT-5.4
$2.50
$15.00
$17.50
OpenAI
Claude Sonnet 4.5
$3.00
$15.00
$18.00
Anthropic
Claude Opus 4.6
$5.00
$25.00
$30.00
Anthropic
GPT-5.4 Pro
$30.00
$180.00
$210.00
OpenAI
Notably, the model consumes quota at three times the standard rate during peak hours, which are defined as 14:00 to 18:00 Beijing Time daily, though a limited-time promotion through April 2026 allows off-peak usage to be billed at a standard 1x rate. Complementing the flagship is the recently debuted GLM-5 Turbo.
While 5.1 is the marathon runner, Turbo is the sprinter, proprietary and optimized for fast inference and tasks like tool use and persistent automation.
At a cost of $1.20 per million input / $4 per million output, it is more expensive than the base GLM-5 but comes in at more affordable than the new GLM-5.1, positioning it as a commercially attractive option for high-speed, supervised agent runs.
The model is also packaged for local deployment, supporting inference frameworks including vLLM, SGLang, and xLLM. Comprehensive deployment instructions are available at the official GitHub repository, allowing developers to run the 754 billion parameter MoE model on their own infrastructure.
For enterprise teams, the model includes advanced reasoning capabilities that can be accessed via a thinking parameter in API requests, allowing the model to show its step-by-step internal reasoning process before providing a final answer.
Benchmarks: a new global standard
The performance data for GLM-5.1 suggests it has leapfrogged several established Western models in coding and engineering tasks.
On SWE-Bench Pro, which evaluates a model's ability to resolve real-world GitHub issues using an instruction prompt and a 200,000 token context window, GLM-5.1 achieved a score of 58.4. For context, this outperforms GPT-5.4 at 57.7, Claude Opus 4.6 at 57.3, and Gemini 3.1 Pro at 54.2.
Beyond standardized coding tests, the model showed significant gains in reasoning and agentic benchmarks. It scored 63.5 on Terminal-Bench 2.0 when evaluated with the Terminus-2 framework and reached 66.5 when paired with the Claude Code harness.
On CyberGym, it achieved a 68.7 score based on a single-run pass over 1,507 tasks, demonstrating a nearly 20-point lead over the previous GLM-5 model. The model also performed strongly on the MCP-Atlas public set with a score of 71.8 and achieved a 70.6 on the T3-Bench.
In the reasoning domain, it scored 31.0 on Humanitys Last Exam, which jumped to 52.3 when the model was allowed to use external tools. On the AIME 2026 math competition benchmark, it reached 95.3, while scoring 86.2 on GPQA-Diamond for expert-level science reasoning.
The most impressive anecdotal benchmark was the Scenario 3 test: building a Linux-style desktop environment from scratch in eight hours.
Unlike previous models that might produce a basic taskbar and a placeholder window before declaring the task complete, GLM-5.1 autonomously filled out a file browser, terminal, text editor, system monitor, and even functional games.
It iteratively polished the styling and interaction logic until it had delivered a visually consistent, functional web application. This serves as a concrete example of what becomes possible when a model is given the time and the capability to keep refining its own work.
Licensing and the open segue
The licensing of these two models tells a larger story about the current state of the global AI market. GLM-5.1 has been released under the MIT License, with its model weights made publicly available on Hugging Face and ModelScope.
This follows the Z.ai historical strategy of using open-source releases to build developer goodwill and ecosystem reach. However, GLM-5 Turbo remains proprietary and closed-source. This reflects a growing trend among leading AI labs toward a hybrid model: using open-source models for broad distribution while keeping execution-optimized variants behind a paywall.
Industry analysts note that this shift arrives amidst a rebalancing in the Chinese market, where heavyweights like Alibaba are also beginning to segment their proprietary work from their open releases.
Z.ai CEO Zhang Peng appears to be navigating this by ensuring that while the flagship's core intelligence is open to the community, the high-speed execution infrastructure remains a revenue-driving asset.
The company is not explicitly promising to open-source GLM-5 Turbo itself, but says the findings will be folded into future open releases. This segmented strategy helps drive adoption while allowing the company to build a sustainable business model around its most commercially relevant work.
Community and user reactions: crushing a week's work
The developer community response to the GLM-5.1 release has been overwhelmingly focused on the model's reliability in production-grade environments.
User reviews suggest a high degree of trust in the model's autonomy.
One developer noted that GLM-5.1 shocked them with how good it is, stating it seems to do what they want more reliably than other models with less reworking of prompts needed. Another developer mentioned that the model's overall workflow from planning to project execution performs excellently, allowing them to confidently entrust it with complex tasks.
Specific case studies from users highlight significant efficiency gains.
A user from Crypto Economy News reported that a task involving preprocessing code, feature selection logic, and hyperparameter tuning solutions, which originally would have taken a week, was completed in just two days. Since getting the GLM Coding plan, other developers have noted being able to operate more freely and focus on core development without worrying about resource shortages hindering progress.
On social media, the launch announcement generated over 46,000 views in its first hour, with users captivated by the eight-hour autonomous claim. The sentiment among early adopters is that Z.ai has successfully moved past the hallucination-heavy era of AI into a period where models can be trusted to optimize themselves through repeated iteration.
The ability to build four applications rapidly through correct prompting and structured planning has been cited by multiple users as a game-changing development for individual developers.
The implications of long-horizon work
The release of GLM-5.1 suggests that the next frontier of AI competition will not be measured in tokens per second, but in autonomous duration.
If a model can work for eight hours without human intervention, it fundamentally changes the software development lifecycle.
However, Z.ai acknowledges that this is only the beginning. Significant challenges remain, such as developing reliable self-evaluation for tasks where no numeric metric exists to optimize against.
Escaping local optima earlier when incremental tuning stops paying off is another major hurdle, as is maintaining coherence over execution traces that span thousands of tool calls.
For now, Z.ai has placed a marker in the sand. With GLM-5.1, they have delivered a model that doesn't just answer questions, but finishes projects. The model is already compatible with a wide range of developer tools including Claude Code, OpenCode, Kilo Code, Roo Code, Cline, and Droid.
For developers and enterprises, the question is no longer, "what can I ask this AI?" but "what can I assign to it for the next eight hours?"
The focus of the industry is clearly shifting toward systems that can reliably execute multi-step work with less supervision. This transition to agentic engineering marks a new phase in the deployment of artificial intelligence within the global economy.
- Gemma 4 opens doors with Apache 2.0, shifting enterprise AI choices.
For the past two years, enterprises evaluating open-weight models have faced an awkward trade-off. Google's Gemma line consistently delivered strong performance, but its custom license — with usage restrictions and terms Google could update at will — pushed many teams toward Mistral or Alibaba's Qwen instead. Legal review added friction. Compliance teams flagged edge cases. And capable as Gemma 3 was, "open" with asterisks isn't the same as open.
Gemma 4 eliminates that friction entirely. Google DeepMind's newest open model family ships under a standard Apache 2.0 license — the same permissive terms used by Qwen, Mistral, Arcee, and most of the open-weight ecosystem.
No custom clauses, no "Harmful Use" carve-outs that required legal interpretation, no restrictions on redistribution or commercial deployment. For enterprise teams that had been waiting for Google to play on the same licensing terms as the rest of the field, the wait is over.
The timing is notable. As some Chinese AI labs (most notably Alibaba’s latest Qwen models, Qwen3.5 Omni and Qwen 3.6 Plus) have begun pulling back from fully open releases for their latest models, Google is moving in the opposite direction — opening up its most capable Gemma release yet while explicitly stating the architecture draws from its commercial Gemini 3 research.
Four models, two tiers: Edge to workstation in a single family
Gemma 4 arrives as four distinct models organized into two deployment tiers. The "workstation" tier includes a 31B-parameter dense model and a 26B A4B Mixture-of-Experts model — both supporting text and image input with 256K-token context windows. The "edge" tier consists of the E2B and E4B, compact models designed for phones, embedded devices, and laptops, supporting text, image, and audio with 128K-token context windows.
The naming convention takes some unpacking. The "E" prefix denotes "effective parameters" — the E2B has 2.3 billion effective parameters but 5.1 billion total, because each decoder layer carries its own small embedding table through a technique Google calls Per-Layer Embeddings (PLE). These tables are large on disk but cheap to compute, which is why the model runs like a 2B while technically weighing more.
The "A" in 26B A4B stands for "active parameters" — only 3.8 billion of the MoE model's 25.2 billion total parameters activate during inference, meaning it delivers roughly 26B-class intelligence with compute costs comparable to a 4B model.
For IT leaders sizing GPU requirements, this translates directly to deployment flexibility. The MoE model can run on consumer-grade GPUs and should appear quickly in tools like Ollama and LM Studio. The 31B dense model requires more headroom — think an NVIDIA H100 or RTX 6000 Pro for unquantized inference — but Google is also shipping Quantization-Aware Training (QAT) checkpoints to maintain quality at lower precision. On Google Cloud, both workstation models can now run in a fully serverless configuration via Cloud Run with NVIDIA RTX Pro 6000 GPUs, spinning down to zero when idle.
The MoE bet: 128 small experts to save on inference costs
The architectural choices inside the 26B A4B model deserve particular attention from teams evaluating inference economics. Rather than following the pattern of recent large MoE models that use a handful of big experts, Google went with 128 small experts, activating eight per token plus one shared always-on expert. The result is a model that benchmarks competitively with dense models in the 27B–31B range while running at roughly the speed of a 4B model during inference.
This is not just a benchmark curiosity — it directly affects serving costs. A model that delivers 27B-class reasoning at 4B-class throughput means fewer GPUs, lower latency, and cheaper per-token inference in production. For organizations running coding assistants, document processing pipelines, or multi-turn agentic workflows, the MoE variant may be the most practical choice in the family.
Both workstation models use a hybrid attention mechanism that interleaves local sliding window attention with full global attention, with the final layer always global. This design enables the 256K context window while keeping memory consumption manageable — an important consideration for teams processing long documents, codebases, or multi-turn agent conversations.
Native multimodality: Vision, audio, and function calling baked in from scratch
Previous generations of open models typically treated multimodality as an add-on. Vision encoders were bolted onto text backbones. Audio required an external ASR pipeline like Whisper. Function calling relied on prompt engineering and hoping the model cooperated. Gemma 4 integrates all of these capabilities at the architecture level.
All four models handle variable aspect-ratio image input with configurable visual token budgets — a meaningful improvement over Gemma 3n's older vision encoder, which struggled with OCR and document understanding. The new encoder supports budgets from 70 to 1,120 tokens per image, letting developers trade off detail against compute depending on the task.
Lower budgets work for classification and captioning; higher budgets handle OCR, document parsing, and fine-grained visual analysis. Multi-image and video input (processed as frame sequences) are supported natively, enabling visual reasoning across multiple documents or screenshots.
The two edge models add native audio processing — automatic speech recognition and speech-to-translated-text, all on-device. The audio encoder has been compressed to 305 million parameters, down from 681 million in Gemma 3n, while the frame duration dropped from 160ms to 40ms for more responsive transcription. For teams building voice-first applications that need to keep data local — think healthcare, field service, or multilingual customer interaction — running ASR, translation, reasoning, and function calling in a single model on a phone or edge device is a genuine architectural simplification.
Function calling is also native across all four models, drawing on research from Google's FunctionGemma release late last year. Unlike previous approaches that relied on instruction-following to coax models into structured tool use, Gemma 4's function calling was trained into the model from the ground up — optimized for multi-turn agentic flows with multiple tools. This shows up in agentic benchmarks, but more importantly, it reduces the prompt engineering overhead that enterprise teams typically invest when building tool-using agents.
Benchmarks in context: Where Gemma 4 lands in a crowded field
The benchmark numbers tell a clear story of generational improvement. The 31B dense model scores 89.2% on AIME 2026 (a rigorous mathematical reasoning test), 80.0% on LiveCodeBench v6, and hits a Codeforces ELO of 2,150 — numbers that would have been frontier-class from proprietary models not long ago. On vision, MMMU Pro reaches 76.9% and MATH-Vision hits 85.6%.
For comparison, Gemma 3 27B scored 20.8% on AIME and 29.1% on LiveCodeBench without thinking mode.
The MoE model tracks closely: 88.3% on AIME 2026, 77.1% on LiveCodeBench, and 82.3% on GPQA Diamond — a graduate-level science reasoning benchmark. The performance gap between the MoE and dense variants is modest given the significant inference cost advantage of the MoE architecture.
The edge models punch above their weight class. The E4B hits 42.5% on AIME 2026 and 52.0% on LiveCodeBench — strong for a model that runs on a T4 GPU. The E2B, smaller still, manages 37.5% and 44.0% respectively. Both significantly outperform Gemma 3 27B (without thinking) on most benchmarks despite being a fraction of the size, thanks to the built-in reasoning capability.
These numbers need to be read against an increasingly competitive open-weight landscape. Qwen 3.5, GLM-5, and Kimi K2.5 all compete aggressively in this parameter range, and the field moves fast. What distinguishes Gemma 4 is less any single benchmark and more the combination: strong reasoning, native multimodality across text, vision, and audio, function calling, 256K context, and a genuinely permissive license — all in a single model family with deployment options from edge devices to cloud serverless.
What enterprise teams should watch next
Google is releasing both pre-trained base models and instruction-tuned variants, which matters for organizations planning to fine-tune for specific domains. The Gemma base models have historically been strong foundations for custom training, and the Apache 2.0 license now removes any ambiguity about whether fine-tuned derivatives can be deployed commercially.
The serverless deployment option via Cloud Run with GPU support is worth watching for teams that need inference capacity that scales to zero. Paying only for actual compute during inference — rather than maintaining always-on GPU instances — could meaningfully change the economics of deploying open models in production, particularly for internal tools and lower-traffic applications.
Google has hinted that this may not be the complete Gemma 4 family, with additional model sizes likely to follow. But the combination available today — workstation-class reasoning models and edge-class multimodal models, all under Apache 2.0, all drawing from Gemini 3 research — represents the most complete open model release Google has shipped. For enterprise teams that had been waiting for Google's open models to compete on licensing terms as well as performance, the evaluation can finally begin without a call to legal first.
- DeepSeek reemerges with powerful AI at a fraction of the cost
The whale has resurfaced.
DeepSeek, the Chinese AI startup offshoot of High-Flyer Capital Management quantitative analysis firm, became a near-overnight sensation globally in January 2025 with the release of its open source R1 model that matched proprietary U.S. giants.
It's been an epoch in AI since then, and while DeepSeek has released several updates to that model and its other V3 series, the international AI and business community has been largely waiting with baited breath for the follow-up to the R1 moment.
Now it's arrived with last night's release of DeepSeek-V4, a 1.6-trillion-parameter Mixture-of-Experts (MoE) model available free under commercially-friendly open source MIT License, which nears — and on some benchmarks, surpasses — the performance of the world’s most advanced closed-source systems at approximately 1/6th the cost over the application programming interface (API).
This release—which DeepSeek AI researcher Deli Chen described on X as a "labor of love" 484 days after the launch of V3—is being hailed as the "second DeepSeek moment".
As Chen noted in his post, "AGI belongs to everyone". It's available now on AI code sharing community Hugging Face and through DeepSeek's API.
Frontier-class AI gets pushed into a lower price band
The most immediate impact of the DeepSeek-V4 launch is economic. The corrected pricing table shows DeepSeek is not pricing its new Pro model at near-zero levels, but it is still pushing high-end model access into a far lower cost tier than the leading U.S. frontier models.
DeepSeek-V4-Pro is priced through its API at $1.74 USD per 1 million input tokens on a cache miss and $3.48 per million output tokens.
That puts a simple one-million-input, one-million-output comparison at $5.22. With cached input, the input price drops to $0.145 per million tokens, bringing that same blended comparison down to $3.625.
That is dramatically cheaper than the current premium pricing from OpenAI and Anthropic. GPT-5.5 is priced at $5.00 per million input tokens and $30.00 per million output tokens, for a combined $35.00 in the same simple comparison.
Claude Opus 4.7 is priced at $5.00 input and $25.00 output, for a combined $30.00.
Model
Input
Output
Total Cost
Source
Grok 4.1 Fast
$0.20
$0.50
$0.70
xAI
MiniMax M2.7
$0.30
$1.20
$1.50
MiniMax
Gemini 3 Flash
$0.50
$3.00
$3.50
Google
Kimi-K2.5
$0.60
$3.00
$3.60
Moonshot
MiMo-V2-Pro (≤256K)
$1.00
$3.00
$4.00
Xiaomi MiMo
GLM-5
$1.00
$3.20
$4.20
Z.ai
GLM-5-Turbo
$1.20
$4.00
$5.20
Z.ai
DeepSeek-V4-Pro
$1.74
$3.48
$5.22
DeepSeek
GLM-5.1
$1.40
$4.40
$5.80
Z.ai
Claude Haiku 4.5
$1.00
$5.00
$6.00
Anthropic
Qwen3-Max
$1.20
$6.00
$7.20
Alibaba Cloud
Gemini 3 Pro
$2.00
$12.00
$14.00
Google
GPT-5.2
$1.75
$14.00
$15.75
OpenAI
GPT-5.4
$2.50
$15.00
$17.50
OpenAI
Claude Sonnet 4.5
$3.00
$15.00
$18.00
Anthropic
Claude Opus 4.7
$5.00
$25.00
$30.00
Anthropic
GPT-5.5
$5.00
$30.00
$35.00
OpenAI
GPT-5.4 Pro
$30.00
$180.00
$210.00
OpenAI
On standard, cache-miss pricing, DeepSeek-V4-Pro comes in at roughly one-seventh the cost of GPT-5.5 and about one-sixth (1/6th) the cost of Claude Opus 4.7.
With cached input, the gap widens: DeepSeek-V4-Pro costs about one-tenth as much as GPT-5.5 and about one-eighth as much as Claude Opus 4.7.
The more extreme near-zero story belongs to DeepSeek-V4-Flash, not the Pro model. Flash is priced at $0.14 per million input tokens on a cache miss and $0.28 per million output tokens, for a combined $0.42.
With cached input, that drops to $0.308. In that case, DeepSeek’s cheaper model is more than 98% below GPT-5.5 and Claude Opus 4.7 in a simple input-plus-output comparison, or nearly 1/100th the cost — though the performance dips significantly.
DeepSeek is compressing advanced model economics into a much lower band, forcing developers and enterprises to revisit the cost-benefit calculation around premium closed models.
For companies running large inference workloads, that price gap can change what is worth automating. Tasks that look too expensive on GPT-5.5 or Claude Opus 4.7 may become economically viable on DeepSeek-V4-Pro, and even more so on DeepSeek-V4-Flash. The launch does not make intelligence free, but it does make the market harder for premium providers to defend on performance alone.
Benchmarking the frontier: DeepSeek-V4-Pro gets close, but GPT-5.5 and Opus 4.7 still lead on most shared tests
DeepSeek-V4-Pro-Max is best understood as a major open-weight leap, not a clean across-the-board defeat of the newest closed frontier systems.
The model’s strongest benchmark claims come from DeepSeek’s own comparison tables, where it is shown against GPT-5.4 xHigh, Claude Opus 4.6 Max and Gemini 3.1 Pro High and bests them on several tests, including Codeforces and Apex Shortlist.
But that is not the same as a head-to-head against OpenAI’s newer GPT-5.5 or Anthropic’s newer Claude Opus 4.7.
Looking only at DeepSeek-V4 versus the latest proprietary models, the picture is more restrained.
On this shared set, GPT-5.5 and Claude Opus 4.7 still lead most categories.
DeepSeek-V4-Pro-Max’s best showing is on BrowseComp, the benchmark measuring agentic AI web browsing prowess (especially highly containerized information), where it scores 83.4%, narrowly behind GPT-5.5 at 84.4% and ahead of Claude Opus 4.7 at 79.3%.
On Terminal-Bench 2.0, DeepSeek scores 67.9%, close to Claude Opus 4.7’s 69.4%, but far behind GPT-5.5’s 82.7%.
Benchmark
DeepSeek-V4-Pro-Max
GPT-5.5
GPT-5.5 Pro, where shown
Claude Opus 4.7
Best result among these
GPQA Diamond
90.1%
93.6%
—
94.2%
Claude Opus 4.7
Humanity’s Last Exam, no tools
37.7%
41.4%
43.1%
46.9%
Claude Opus 4.7
Humanity’s Last Exam, with tools
48.2%
52.2%
57.2%
54.7%
GPT-5.5 Pro
Terminal-Bench 2.0
67.9%
82.7%
—
69.4%
GPT-5.5
SWE-Bench Pro / SWE Pro
55.4%
58.6%
—
64.3%
Claude Opus 4.7
BrowseComp
83.4%
84.4%
90.1%
79.3%
GPT-5.5 Pro
MCP Atlas / MCPAtlas Public
73.6%
75.3%
—
79.1%
Claude Opus 4.7
The shared academic-reasoning results favor the closed models: On GPQA Diamond, DeepSeek-V4-Pro-Max scores 90.1%, while GPT-5.5 reaches 93.6% and Claude Opus 4.7 reaches 94.2%.
On Humanity’s Last Exam without tools, DeepSeek scores 37.7%, behind GPT-5.5 at 41.4%, GPT-5.5 Pro at 43.1% and Claude Opus 4.7 at 46.9%. With tools enabled, DeepSeek rises to 48.2%, but still trails GPT-5.5 at 52.2%, GPT-5.5 Pro at 57.2% and Claude Opus 4.7 at 54.7%.
The agentic and software-engineering results are more mixed, but they still show DeepSeek-V4-Pro-Max trailing GPT-5.5 and Opus 4.7.
On Terminal-Bench 2.0, DeepSeek’s 67.9% is competitive with Claude Opus 4.7’s 69.4%, but GPT-5.5 is much higher at 82.7%.
On SWE-Bench Pro, DeepSeek’s 55.4% trails GPT-5.5 at 58.6% and Claude Opus 4.7 at 64.3%. On MCP Atlas, DeepSeek’s 73.6% is slightly behind GPT-5.5 at 75.3% and Claude Opus 4.7 at 79.1%.
BrowseComp is the standout: DeepSeek’s 83.4% beats Claude Opus 4.7’s 79.3% and nearly matches GPT-5.5’s 84.4%, though GPT-5.5 Pro’s 90.1% remains well ahead.
So ultimately, DeepSeek-V4-Pro-Max does not appear to dethrone GPT-5.5 or Claude Opus 4.7 on the benchmarks that can be directly compared across the companies’ published tables. But it gets close enough on several of them — especially BrowseComp, Terminal-Bench 2.0 and MCP Atlas — that its much lower API pricing becomes the headline.
In practical terms, DeepSeek does not need to win every leaderboard row to matter. If it can deliver near-frontier performance on many enterprise-relevant agent and reasoning tasks at roughly one-sixth to one-seventh the standard API cost of GPT-5.5 or Claude Opus 4.7, it still forces a major rethink of the economics of advanced AI deployment.
DeepSeek-V4-Pro-Max is clearly the strongest open-weight model in the field right now, and it is unusually close to frontier closed systems on several practical benchmarks.
While GPT-5.5 and Claude Opus 4.7 still retain the lead in most direct head-to-head comparisons across the company's benchmark charts, DeepSeek V4 Pro gets close while being dramatically cheaper and openly available.
A big jump from DeepSeek V3.2
To understand the magnitude of this release, one must look at the performance gains of the base models. DeepSeek-V4-Pro-Base represents a significant advancement over the previous generation, DeepSeek-V3.2-Base. In World Knowledge, V4-Pro-Base achieved 90.1 on MMLU (5-shot) compared to V3.2’s 87.8, and a massive jump on MMLU-Pro from 65.5 to 73.5.
The improvement in high-level reasoning and verified facts is even more pronounced: on SuperGPQA, V4-Pro-Base reached 53.9 compared to V3.2's 45.0, and on the FACTS Parametric benchmark, it more than doubled its predecessor's performance, jumping from 27.1 to 62.6. Simple-QA verified scores also saw a dramatic rise from 28.3 to 55.2.
The Long Context capabilities have also been refined. On LongBench-V2, V4-Pro-Base scored 51.5, significantly outpacing the 40.2 achieved by V3.2-Base. In Code and Math, V4-Pro-Base reached 76.8 on HumanEval (Pass@1), up from 62.8 on V3.2-Base.
These numbers underscore that DeepSeek has not just optimized for inference cost, but has fundamentally improved the intelligence density of its base architecture. The efficiency story is equally compelling for the Flash variant. DeepSeek-V4-Flash-Base, despite utilizing a substantially smaller number of parameters, outperforms the larger V3.2-Base across wide benchmarks, particularly in long-context scenarios.
A new information 'traffic controller,' Manifold-Constrained Hyper-Connections (mHC)
DeepSeek’s ability to offer these prices and performance figures is rooted in radical architectural innovations detailed in its technical report also released today, "Towards Highly Efficient Million-Token Context Intelligence."
The standout technical achievement of V4 is its native one-million-token context window. Historically, maintaining such a large context required massive memory (the key values or KV cache).
DeepSeek solved this by introducing a Hybrid Attention Architecture that combines Compressed Sparse Attention (CSA) to reduce initial token dimensionality and Heavily Compressed Attention (HCA) to aggressively compress the memory footprint for long-range dependencies.
In practice, the V4-Pro model requires only 10% of the KV cache and 27% of the single-token inference FLOPs compared to its predecessor, the DeepSeek-V3.2, even when operating at a 1M token context.
To stabilize a network of 1.6 trillion parameters, DeepSeek moved beyond traditional residual connections. The company's researchers incorporated Manifold-Constrained Hyper-Connections (mHC) to strengthen signal propagation across layers while preserving the model’s expressivity.
mHC allows an AI to have a much wider flow of information (so it can learn more complex things) without the risk of the model becoming unstable or "breaking" during its training. It’s like giving a city a 10-lane highway but adding a perfect AI traffic controller to ensure no one ever hits the brakes.
This is paired with the Muon optimizer, which allowed the team to achieve faster convergence and greater training stability during the pre-training on more than 32T diverse and high-quality tokens.
This pre-training data was refined to remove hatched auto-generated content, mitigating the risk of model collapse and prioritizing unique academic values. The model’s 1.6T parameters utilize a Mixture-of-Experts (MoE) design where only 49B parameters are activated per token, further driving down compute requirements.
Training the mixture-of-experts (MoE) to work as a whole
DeepSeek-V4 was not simply trained; it was "cultivated" through a unique two-stage paradigm.
First, through Independent Expert Cultivation, domain-specific experts were trained through Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) using the GRPO (Group Relative Policy Optimization) algorithm. This allowed each expert to master specialized skills like mathematical reasoning or codebase analysis.
Second, Unified Model Consolidation integrated these distinct proficiencies into a single model via on-policy distillation, where the unified model acts as the student learning to optimize reverse KL loss with teacher models. This distillation process ensures that the model preserves the specialized capabilities of each expert while operating as a cohesive whole.
The model’s reasoning capabilities are further segmented into three increasing "effort" modes.
The "Non-think" mode provides fast, intuitive responses for routine tasks.
"Think High" provides conscious logical analysis for complex problem-solving.
Finally, "Think Max" pushes the boundaries of model reasoning, bridging the gap with frontier models on complex reasoning and agentic tasks. This flexibility allows users to match the compute effort to the difficulty of the task, further enhancing cost-efficiency.
Breaking the Nvidia GPU stranglehold with local Chinese Huawei Ascend NPUs
While the model weights are the headline, the software stack released alongside them is arguably more important for the future of "Sovereign AI."
Analyst Rui Ma highlighted a single sentence from the release as the most critical: DeepSeek validated their fine-grained Expert Parallelism (EP) scheme on Huawei Ascend NPUs (neural processing units).
By achieving a 1.50x to 1.73x speedup on non-Nvidia GPU platforms, DeepSeek has provided a blueprint for high-performance AI deployment that is resilient to Western GPU supply chains and export controls.
However, it's important to note that DeepSeek still claims it used officially licensed, legal Nvidia GPUs for DeepSeek V4's training, in addition to the Huawei NPUs.
DeepSeek has also open-sourced the MegaMoE mega-kernel as a component of its DeepGEMM library. This CUDA-based implementation delivers up to a 1.96x speedup for latency-sensitive tasks like RL rollouts and high-speed agent serving.
This move ensures that developers can run these massive models with extreme efficiency on existing hardware, further cementing DeepSeek’s role as the primary driver of open-source AI infrastructure.
The technical report emphasizes that these optimizations are crucial for supporting a standard 1M context across all official services.
Licensing and local deployment
DeepSeek-V4 is released under the MIT License, the most permissive framework in the industry. This allows developers to use, copy, modify, and distribute the weights for commercial purposes without royalties—a stark contrast to the "restricted" open-weight licenses favored by other companies.
For local deployment, DeepSeek recommends setting sampling parameters to temperature = 1.0 and top_p = 1.0. For those utilizing the "Think Max" reasoning mode, the team suggests setting the context window to at least 384K tokens to avoid truncating the model's internal reasoning chains.
The release includes a dedicated encoding folder with Python scripts demonstrating how to encode messages in OpenAI-compatible format and parse the model's output, including reasoning content.
DeepSeek-V4 is also seamlessly integrated with leading AI agents like Claude Code, OpenClaw, and OpenCode. This native integration underscores its role as a bedrock for developer tools, providing an open-source alternative to the proprietary ecosystems of major cloud providers.
Community reactions and what comes next
The community reaction has been one of shock and validation. Hugging Face officially welcomed the "whale" back, stating that the era of cost-effective 1M context length has arrived.
Industry experts noted that the "second DeepSeek moment" has effectively reset the developmental trajectory of the entire field, placing massive pressure on closed-source providers like OpenAI and Anthropic to justify their premiums.
AI evaluation firm Vals AI noted that DeepSeek-V4 is now the "#1 open-weight model on our Vibe Code Benchmark, and it’s not close".
DeepSeek is moving quickly to retire its older architectures. The company announced that the legacy deepseek-chat and deepseek-reasoner endpoints will be fully retired on July 24, 2026. All traffic is currently being rerouted to the V4-Flash architecture, signifying a total transition to the million-token standard.
DeepSeek-V4 is more than just a new model; it is a challenge to the status quo. By proving that architectural innovation can substitute for raw compute-maximalism, DeepSeek has made the highest levels of AI intelligence accessible to the global developer community at a far lower cost — something that could benefit the globe, even at a time when lawmakers and leaders in Washington, D.C. are raising concerns about Chinese labs "distilling" from U.S. proprietary giants to train open source models, and fears of said open source or jailbroken proprietary models being used to create weapons and commit terror.
The truth is, while all of these are potential risks — as they were and have been with prior technologies that broadened information access, like search and the internet itself — the benefits seem far outweigh them, and DeepSeek's quest to keep frontier AI models open is of benefit to the entire planet of potential AI users, especially enterprises looking to adopt the cutting-edge at the lowest possible cost.
