Open-Weight vs Closed-Source AI Models 2026: Gap Analysis
Q2 2026 gap analysis between open-weight and closed-source frontier models — capability parity, cost economics, and the agency deployment decision tree.
Chinese Provider Share
Open-Weight Frontier Models
Largest Open-Weight Context
M2.7 vs Opus Cost Gap
Key Takeaways
Open-weight models used to be the budget option. In Q2 2026, they are the volume leaders. Xiaomi's MiMo V2 Pro processes more tokens than any model on OpenRouter — and you can download it.
The narrative that closed-source frontier models are the only serious production option is out of date. OpenRouter traffic tells a different story: Chinese open-weight providers combined now account for more than 45% of all tokens flowing through the aggregator, up from under 2% a year ago. That inversion changes how agencies should think about model selection, total cost of ownership, and vendor lock-in risk.
This analysis is structured as a gap comparison. Rather than ranking models, it walks through the specific capability dimensions where open-weight has caught up, where it has not, and where the economics force a rethink of the default choice. The goal is a decision framework that an agency can actually apply to a client workload on Monday morning.
Data snapshot: Benchmark and usage figures reflect publicly reported data as of early April 2026. Model releases continue at pace and the landscape shifts monthly. See our OpenRouter April 2026 rankings for the underlying traffic data.
The Q2 2026 Inversion: Open Weights Win Volume
A year ago, the assumption was that frontier capability required closed weights. The reasoning was straightforward: training a frontier model cost nine figures, and whoever paid that bill protected the outcome. In Q2 2026, that assumption still holds for the very top of the reasoning leaderboard, but it has broken everywhere else.
The single cleanest data point: on OpenRouter, Xiaomi alone accounts for 21.1% of all weekly tokens, which is roughly three times OpenAI's 7.5%. Alibaba's Qwen family adds another 13.9%. MiniMax, Zhipu, DeepSeek, and StepFun together add another 24.6%. Anthropic sits at 10.9% and Google at 11.3%. The majority of developer traffic through the leading model aggregator now flows through Chinese providers, most of them shipping open weights.
The Open-Weight Frontier Cohort
Seven open-weight models currently sit in the frontier tier as of April 2026:
- MiMo V2 Pro (Xiaomi, March 18): 1T+ parameter MoE with 42B active, 1M context, $1 input and $3 output per million tokens. Currently #1 on OpenRouter by a 3x margin.
- MiniMax M2.7 (March 18): Self-evolving MoE with 10B active parameters, 205K context, $0.30 input and $1.20 output per million. Hits 56.22% on SWE-Pro.
- Step 3.5 Flash (StepFun, February 2): 196B MoE with 11B active, 262K context, free tier available on OpenRouter. Currently #3 by volume.
- Nemotron 3 Super 120B (NVIDIA, March 10-11): 120B total with 12B active, 60.47% on SWE-Bench Verified, ships under NVIDIA's Open Model License with clean commercial terms.
- DeepSeek V3.2 (December 2025): 685B MoE, strong reasoning, IMO gold medal performance, long-context capable. DeepSeek V4 is expected but not released as of April 2026.
- Kimi K2.5 (Moonshot): 1T MoE with Agent Swarm architecture, 262K context, $0.38 input and $1.72 output per million. Backbone of Cursor Composer 2.
- Gemma 4 31B Dense (Google, April 2): Apache 2.0 licensed, built from Gemini 3 internals, ranked #3 among open models globally with 84.3% on GPQA Diamond.
The Closed-Source Frontier Cohort
On the closed-source side, six models dominate production agency-class workloads:
- Claude Opus 4.6 (Anthropic): Still the top agentic coding model as of early April 2026, with 80%+ on SWE-bench Verified. Dominates enterprise Anthropic spend at roughly $25.1M per month in reported consumption.
- Claude Sonnet 4.6 (Anthropic, February 17): Near-Opus quality at $3/$15 per million. Default on claude.ai Free and Pro tiers.
- GPT-5.4 (OpenAI, March 5): SOTA on computer use, 83% on GDPval, 1M context in Codex mode. $2.50/$15 per million on API.
- Gemini 3.1 Pro (Google): 2M context, unified multimodal, 77.1% on ARC-AGI-2. $2/$12 per million.
- Qwen 3.6 Plus (Alibaba, April 2): 1M context, 65K output, always-on chain-of-thought. Closed weights despite Alibaba's open-weight history. Free during preview.
- Qwen 3.5-Omni (Alibaba, March 30): Native omnimodal covering text, image, audio, video. 256K context, 113 languages. Ships mostly closed-source.
Not every Chinese model is open. Alibaba ships Qwen 3.6 Plus and Qwen 3.5-Omni as closed weights even though earlier Qwen versions were open. Treating nationality as a proxy for licensing is a common and expensive mistake. Our AI Digital Transformation practice handles this licensing review for client procurement.
Capability Dimensions Compared
"Open versus closed" only makes sense when you decompose it into the actual capabilities that matter for a workload. Seven dimensions shape the picture:
Closed-source leads. Opus 4.6 and GPT-5.4 Pro sit at the top of GPQA Diamond and FrontierMath. Gemma 4 31B Dense and DeepSeek V3.2 are closest open-weight challengers.
Gap has effectively closed. MiMo V2 Pro leads OpenRouter coding tokens. MiniMax M2.7 hits 56.22% on SWE-Pro at 50x less cost than Opus. Closed-source holds SOTA but not by much.
Closed leads on reliability; open leads on volume. Opus 4.6 and GPT-5.4 top MCP-Atlas. MiMo V2 Pro, Qwen 3.6 Plus, and MiniMax M2.7 together account for over 27% of tool calls on OpenRouter.
Clear closed-source lead. GPT-5.4 native computer use at 75% OSWorld. Gemini 3.1 Pro unified multimodal. MiMo V2 Omni and Qwen 3.5-Omni are the closest open/closed-Chinese response.
Open-weight matches and sometimes beats. MiMo V2 Pro and Qwen 3.6 Plus both ship 1M context. Gemini 3.1 Pro leads at 2M but Opus 4.6's 1M is still beta-tier.
Closed-source has a structural advantage. Anthropic's Project Glasswing and OpenAI's Aardvark framework apply formal guardrails. Open weights can be un-aligned by any downstream fine-tune.
The Seventh Dimension: Instruction Following
On literal instruction following, the newest closed-source releases have taken a decisive lead over open-weight. Opus 4.6 followed instructions more loosely than some open-weight models; the April 2026 release of Opus 4.7 inverts that, executing prompts as written rather than generalizing. Open-weight models tend to preserve the older "helpful generalization" behavior, which is either a feature or a bug depending on the workload.
Reasoning Gap: Where Closed-Source Still Wins
On hard reasoning benchmarks, closed-source models retain a meaningful lead. The gap is not the 30+ points it was in 2024; in most categories it now sits in the 3 to 8 percentage point range. But for workloads where getting the right answer matters more than getting a cheap answer, that gap is load-bearing.
Where the Gap Is Largest
- Humanity's Last Exam with tools: Closed-source leads by roughly 5-10 points. GPT-5.4 Pro at 58.7% and the Mythos-class ceiling at 64.7% are not matched by any open-weight release.
- FrontierMath Tier 4: GPT-5.4 Pro at 38.0% with tools. Open-weight releases, including DeepSeek V3.2's IMO gold medal performance, still sit notably below this.
- ARC-AGI-2 Verified: Gemini 3.1 Pro at 77.1% and GPT-5.4 Pro at 83.3%. Open-weight releases have not publicly posted competitive numbers here.
- Deep Think / Extended Reasoning: Gemini 3.1 Pro's Deep Think mode and Opus 4.6's adaptive thinking at max effort produce measurably better results on hard-problem traces than open-weight equivalents running at similar token budgets.
Where the Gap Is Narrowest
On GPQA Diamond, Gemma 4 31B Dense scores 84.3% against closed-source leaders in the 94% range, which is a gap but not a chasm. AIME 89.2% from Gemma 4 is competitive. MiniMax M2.7 and DeepSeek V3.2 both post respectable results on reasoning benchmarks that would have been closed-source exclusives 18 months ago.
The practical read: for reasoning-critical workloads, default to Opus 4.6, GPT-5.4 Pro, or Gemini 3.1 Pro Deep Think. For reasoning-adjacent workloads where the task is complex but not frontier-hard, the open-weight frontier is now genuinely competitive.
Coding Gap: Where Open-Weight Has Closed
Coding is the clearest case of open-weight parity. On OpenRouter, MiMo V2 Pro and Qwen 3.6 Plus combined move approximately 49% of all coding tokens. MiMo V2 Pro alone is 25.5% of coding tokens, more than six times Anthropic's share. That is not a future prediction — it is the current developer reality.
- MiMo V2 Pro: 1T+ parameter MoE, 1M context, $1/$3 per million. Leads OpenRouter coding volume by a 6x margin over Anthropic.
- MiniMax M2.7: 56.22% SWE-Pro, $0.30/$1.20 per million, roughly 50x cheaper than Opus on output tokens.
- DeepSeek V3.2: 685B params, strong long-context coding performance, industry workhorse for cost-sensitive pipelines.
- Kimi K2.5: Powers Cursor Composer 2 at 73.7% on SWE-bench Multilingual, proof that open weights now back shipping IDE products.
- Nemotron 3 Super 120B: 60.47% on SWE-Bench Verified under NVIDIA's commercial-friendly open license.
Closed-source still leads on the very top of the coding leaderboard. GPT-5.3-Codex holds 78.2% on SWE-bench Pro Public. Opus 4.6 sits above 80% on SWE-bench Verified. GPT-5.4 posts 57.7% on SWE-Bench Pro, and Opus 4.7 (released mid-April 2026) jumps to 64.3% SWE-Pro and 87.6% SWE-bench Verified. For the hardest multi-file refactors and long-horizon agentic coding, closed-source is still the safer default.
But for the bulk of coding workloads — autocomplete, PR review, bug fixing, small feature work, test generation — MiMo V2 Pro and Qwen 3.6 Plus are doing the work that developers have chosen them to do. For deeper analysis of the specific cost-per-token tradeoff, see our LLM API pricing index for Q2 2026.
Multimodal Gap: Still a Closed-Source Story
Multimodal is the category where the closed/open gap remains widest in Q2 2026. Unified text-image-audio-video reasoning requires training corpora, infrastructure, and alignment work that open-weight releases have not yet matched at the top end.
Closed-Source Leaders
- GPT-5.4: Native computer use at 75% on OSWorld-Verified, which exceeds the 72.4% human baseline. Best in class for agentic UI automation over real screenshots.
- Gemini 3.1 Pro: Unified multimodal architecture with 2M context, strongest video understanding on the market, and 113+ language coverage for audio.
- Qwen 3.5-Omni (Alibaba, closed): Native omnimodal across text, image, audio, video, with 256K context and 113 languages. Competes directly with Gemini on multimodal coverage.
- Claude Opus 4.7 (April 16): 2,576-pixel image processing, over 3x the resolution of prior Claude models. Strong on dense UI screenshots and technical diagrams.
Open-Weight Contenders
- MiMo V2 Omni (Xiaomi, March 18): Omnimodal across image, video, audio with a unified architecture and 262K context. Strongest open-weight multimodal release to date.
- Gemma 4 multimodal variants (Google, April 2): Apache 2.0 licensed, built from Gemini 3 internals. Competitive on image but not on video.
- Phi-4-reasoning-vision-15B (Microsoft, March 4): Compact multimodal reasoning for edge deployment rather than frontier comparison.
The practical read: if the workload requires video understanding, native computer use, or unified audio-video reasoning, closed source is the default. For image-heavy work that does not require video, MiMo V2 Omni is the first open-weight release that is a serious contender.
Hosting Cost Economics
The open/closed decision eventually becomes an infrastructure decision. Three deployment paths exist for any frontier open- weight model: closed-source API, serverless open-weight endpoint (DeepInfra, Together, Fireworks, Groq, Cerebras), or self-hosted GPU inference. The economics shift dramatically between them.
Closed-Source API Pricing
- Claude Opus 4.6: $5 input / $25 output per million tokens.
- Claude Sonnet 4.6: $3 / $15 per million.
- GPT-5.4: $2.50 / $15 per million.
- Gemini 3.1 Pro: $2 / $12 per million.
Open-Weight Serverless API Pricing
- MiMo V2 Pro (Xiaomi): $1 / $3 per million, 1.04M context.
- MiniMax M2.7: $0.30 / $1.20 per million, 205K context.
- Kimi K2.5: $0.38 / $1.72 per million, 262K context.
- Step 3.5 Flash: $0.10 / $0.30 per million or free tier.
Self-Hosted GPU Inference
Self-hosting on H100 or B200 hardware looks appealing on paper but the economics only work at scale. A reserved 8xH100 node on a 1-year contract runs approximately $16-$20 per hour. For a 100B MoE model at reasonable utilization, that maps to roughly $0.50 to $1.00 per million output tokens, well below MiMo V2 Pro's serverless $3 output price. B200 nodes drop that further.
The utilization trap: Self-hosted economics only beat serverless open-weight endpoints when the GPU fleet sustains above roughly 60% utilization around the clock. Below about 100 million tokens per month in steady-state traffic, serverless from DeepInfra, Together, or Fireworks wins on total cost once idle capacity, failover, monitoring, and ops headcount are factored in.
For most agency deployments, the right answer is serverless open-weight plus closed-source API, not self-hosted inference. Self-hosting makes sense for data-residency requirements, compliance isolation, or truly high-volume workloads (500M+ tokens per month sustained).
Agency Deployment Decision Tree
Six questions, in order. The answers route to a concrete model recommendation without needing to benchmark from scratch on every project.
1. Does the workload require frontier reasoning?
If the tasks involve hard math, scientific reasoning, multi-step legal analysis, or anything that would benefit from Deep Think / max-effort reasoning, route to closed source.
Default choice: Claude Opus 4.6/4.7, GPT-5.4 Pro, or Gemini 3.1 Pro Deep Think.
2. Does the workload require native computer use or video?
Computer-use agents, video understanding, and unified audio-video reasoning are closed-source-only today at the frontier tier.
Default choice: GPT-5.4 for computer use, Gemini 3.1 Pro for video.
3. Is the workload coding or dev-tool-heavy?
For coding agencies handling autocomplete, bug fixing, PR review, small feature work, and test generation at volume, open-weight now has the better cost-per-outcome. The gap on frontier coding benchmarks is too narrow to justify the premium for most work.
Default choice: MiMo V2 Pro for volume coding, Kimi K2.5 for IDE integration, Opus 4.6 reserved for multi-file refactors and long-horizon agents.
4. Do client constraints rule out Chinese-origin models?
US federal, defense-adjacent, financial, and some healthcare clients require non-Chinese-origin models in production. If this applies, MiMo, MiniMax, DeepSeek, Kimi, Step, Qwen, and GLM are all out, regardless of licensing.
Default choice: Nemotron 3 Super 120B (open weights), Gemma 4 (open), or closed-source from Anthropic/OpenAI/Google.
5. What is the expected monthly token volume?
Under 10M tokens per month, any choice works and closed-source cost is negligible. 10M to 100M, serverless open-weight endpoints produce meaningful savings. Above 100M sustained, self-hosted open-weight becomes viable. Above 500M, it is often the right answer.
Default choice: Match deployment complexity to steady-state volume, not peak.
6. What are the data residency requirements?
EU data residency, HIPAA isolation, or any client requirement for on-prem inference forces self-hosting. At that point, open-weight is the only option and model selection comes down to fit and license terms.
Default choice: Nemotron 3 Super 120B, Gemma 4, or DeepSeek V3.2 on hardware you control, with a legal review of each license before deployment.
Need help running this tree on a live workload? Our Analytics & Insights team maps model selection against your actual traffic mix and cost curves before rollout.
Risk and Governance Considerations
Capability and cost are only half of the decision. Governance risk is where open-weight deployments typically surprise agencies that treated the choice as purely technical.
Licensing
"Open weight" does not mean "Apache 2.0." Most of the Chinese open-weight frontier ships under bespoke licenses with commercial thresholds, ethical-use clauses, and in some cases jurisdictional restrictions. NVIDIA's Nemotron 3 Super 120B is one of the few exceptions among the top-tier open-weight set, shipping under the NVIDIA Open Model License with clean commercial terms. Gemma 4 is Apache 2.0. Everything else requires a per-license legal review.
Supply Chain
Once weights ship, downstream fine-tunes can strip safety training, add targeted behaviors, or embed backdoors. Agencies deploying open-weight models in production should source weights from the original publisher or a verified mirror, pin checksums, and treat any fine-tune as requiring re-evaluation of the safety baseline. Anthropic's Project Glasswing and OpenAI's Aardvark framework handle this upstream for closed models.
Export Controls and Geopolitics
US export controls on advanced semiconductors shipped to Chinese AI labs continue to tighten. That affects upstream training economics and, indirectly, the release cadence of Chinese frontier models. For published weights, the practical risk is not weights being withdrawn (they cannot be un-published) but future versions becoming unavailable or more expensive, and procurement classification shifting for some clients.
Regulated-Industry Considerations
HIPAA, PCI, SOC 2 Type II, and EU AI Act alignment all work better with closed-source vendors that publish formal compliance documentation. Open-weight deployments inherit compliance responsibility directly. For agencies serving regulated clients, this alone can make closed source the right default regardless of cost. For governance-focused CRM and data workflows, see our CRM & Automation practice.
Model-by-Model Comparison
Ten models across six dimensions: licensing posture, context window, published pricing, coding benchmark, modalities, and primary use case. Figures reflect publicly reported data as of early April 2026.
| Model | License | Context | Price (in/out per 1M) | Coding Benchmark | Modalities |
|---|---|---|---|---|---|
| Claude Opus 4.6 | Closed | 1M beta | $5 / $25 | 80%+ SWE-Verified | Text, image, code |
| Claude Sonnet 4.6 | Closed | 1M beta | $3 / $15 | 79.6% SWE-Verified | Text, image, code |
| GPT-5.4 | Closed | 1M (Codex) | $2.50 / $15 | 57.7% SWE-Pro | Text, image, computer use |
| Gemini 3.1 Pro | Closed | 2M | $2 / $12 | Not publicly posted | Text, image, audio, video |
| Qwen 3.6 Plus | Closed | 1M | Free (preview) | 23.5% of OR coding tokens | Text, code |
| MiMo V2 Pro | Open (bespoke) | 1.04M | $1 / $3 | 25.5% of OR coding tokens | Text, code |
| MiniMax M2.7 | Open (bespoke) | 205K | $0.30 / $1.20 | 56.22% SWE-Pro | Text, code |
| DeepSeek V3.2 | Open (bespoke) | Long-context | Low (varies) | IMO gold-medal reasoning | Text, code |
| Kimi K2.5 | Open (bespoke) | 262K | $0.38 / $1.72 | 73.7% SWE-Multilingual (Composer 2) | Text, code |
| Nemotron 3 Super 120B | Open (NVIDIA OML) | 262K | Free tier / self-host | 60.47% SWE-Verified | Text, code |
Comparison date: April 2026. AI model releases land monthly; verify current pricing, benchmark scores, and license terms before making production decisions. Our Chinese AI market share report tracks ongoing changes.
Conclusion
The open-weight versus closed-source framing is too binary for Q2 2026. The useful split is capability dimension by capability dimension. Reasoning, multimodal, and computer use still belong to closed source. Volume coding, cost-sensitive pipelines, and long-context workloads now belong to open weight. Almost every agency ends up running both.
The practical discipline is to stop defaulting. Run the decision tree against each workload, cost it at realistic volume, and pay attention to licensing and residency requirements before shipping. MiMo V2 Pro at 4.79 trillion tokens per week is the clearest signal that the old default of "just call Claude" no longer holds. The cheapest model that meets the capability bar is the right choice for the majority of work. For the work that actually needs frontier capability, closed source is still earning the premium.
- Build a two-lane architecture: closed-source for frontier reasoning and multimodal, open-weight for volume coding and cost-sensitive pipelines.
- Serverless open-weight endpoints beat self-hosting below 100M tokens per month in steady-state traffic.
- Read every weight license before procurement sign-off. Apache 2.0 is the exception, not the rule, at the open- weight frontier.
- For regulated industries, default to closed source unless compliance documentation is explicitly in scope for the open-weight deployment.
- Revisit the model mix quarterly. Q2 2026 is an inversion point, not an endpoint.
Building a Multi-Model AI Strategy?
Open-weight and closed-source frontier models each earn their place in a modern production stack. We help agencies and enterprises map workloads to the right model, plan licensing and governance, and ship multi-model pipelines that balance capability, cost, and compliance.
Frequently Asked Questions
Related Guides
Continue exploring open-weight and frontier AI model analysis