GPT-4 equivalent performance cost $20 per million tokens in late 2022. Today, the same capability runs closer to $0.40. That collapse is not the whole story, though. Total AI spend for companies in production has gone up, not down, over the same period. The gap between open-weight and closed frontier models is the reason why - and it is now narrow enough that the choice deserves a deliberate answer, not a default.
The concrete trigger for revisiting this: Alibaba launched Qwen3-Coder on July 23, 2025 - its most advanced agentic AI coding model to date, designed for high-performance software development and complex coding workflows. It landed the same week as this post, which makes it a useful lens. The Qwen3 family is the sharpest illustration available of how fast the open-weight tier has moved, what it genuinely offers, and where it still falls short.
What "open weight" actually means in 2026
Open weight means the model weights are publicly released and can be downloaded, self-hosted, fine-tuned, and quantized. Llama 4, DeepSeek V3, Qwen3, Kimi K2, and GLM-4.6 all fall here. Open weight is the more precise term since licenses vary and not every model ships full training data or code, but the key point is that the weights exist and you can run them yourself.
Closed, managed models such as Claude Sonnet 5, GPT-5.5, and Gemini 3.1 Pro are reached only through a vendor's hosted API, billed per token, with the model itself never leaving the vendor's servers.
The practical consequence: proprietary models deliver peak capability and require no infrastructure - but you pay per token, send your data to a third party, and depend on the provider's roadmap and uptime. Open-weight flips that trade: you can download and run the model on your own hardware, giving you control, privacy, and zero per-token API cost - but you take on the cost and complexity of running them.
The nuance most teams miss is that "open weight" does not require self-hosting. Third-party inference providers serve Qwen3 and Llama 4 at rates that reflect actual GPU cost rather than a platform premium. Qwen3-32B and Llama 4 Scout run at roughly $0.78-$0.83 per million tokens on comparable hardware.
Claude Sonnet 4.6 - a model that outperforms GPT-4 on every benchmark that existed when GPT-4 launched - costs $3 per million input tokens. That 3-4x spread is the number to hold in your head.
Why the Qwen3 family changed the calculus
Announced April 29, 2025, Qwen3's flagship model, Qwen3-235B-A22B, achieves competitive results against DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro on coding, math, and general benchmarks. That is not incremental - it is the first open-weight family that does not require an asterisk on frontier comparisons.
Eight models ship across two architectural families: six dense models from 0.6B to 32B parameters, and two mixture-of-experts models - Qwen3-30B-A3B (30B total, 3B active) and Qwen3-235B-A22B (235B total, 22B active). The MoE numbers matter. A model with 30B total parameters but only 3B active per token costs a fraction of a 30B dense model to serve. The small MoE model, Qwen3-30B-A3B, outperforms QwQ-32B with just one-tenth of the activated parameters.
The second genuinely new thing is the hybrid reasoning toggle. Qwen3 introduces a hybrid thinking/non-thinking mode that you control with a single parameter. That design choice matters more than it sounds: you are not choosing between a reasoning model and a chat model. You are running one model and deciding per-request whether to pay the latency and token cost of chain-of-thought reasoning.
Qwen3 also introduces a "thinking budget" mechanism, allowing you to set a maximum token usage during reasoning - which helps control the depth of reasoning and manage computational resource consumption.
The coding extension, Qwen3-Coder, went further still. Released July 2025 with 480B total parameters and 35B active, it was trained on 7.5 trillion tokens with 70% code focus and achieves state-of-the-art results among open models on SWE-Bench Verified, matching Claude Sonnet 4 on agentic coding benchmarks. For reference: Qwen3-Coder scores 67.0% on SWE-Bench Verified, rising to 69.6% with 500 turns.
Claude Sonnet 4 lands at 70.4% at 500 turns. One point separates them - at a fraction of the API cost.
Where closed frontier APIs still win
Honest accounting matters here. Multimodal is the category where the closed/open gap remains widest. 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.
GPT-5.4 leads on computer use, scoring 83% on GDPval with a 1M context in Codex mode, priced at $2.50/$15 per million input/output tokens. No open-weight model currently competes on native UI automation over live screenshots.
The other gap is operational: hidden self-hosting costs frequently surprise organizations - ML engineering for optimization (quantization, sharding, inference containers), infrastructure management, monitoring and observability, and ongoing performance tuning all add up.
Self-hosted breakeven requires 50%+ GPU utilization for 7B models and only 10%+ for 13B models
- meaning small models need very consistent traffic to justify the infrastructure, while larger models break even more easily.
A useful frame:
| Dimension | Open weight | Closed frontier |
|---|---|---|
| Token cost (text/code) | $0.40-$1.50 / M | $2.50-$15 / M |
| Data residency | Full control | Vendor's servers |
| Multimodal capability | Behind, improving | Leading |
| Computer use | Not competitive | GPT-5.4 leads |
| Ops burden | High (self-host) or moderate (third-party API) | Zero |
| Model updates | You control the upgrade path | Vendor controls |
| Fine-tuning | Freely | Restricted or expensive |
| Reasoning toggle | Per-request (Qwen3) | Separate model tier |
The practical answer for most teams: it is a blend. Use a proprietary frontier model where peak quality matters - complex coding, customer-facing answers, hard reasoning - and an open or value model for high-volume, cost-sensitive, or privacy-sensitive work.
Beagle in action#eng-ops, 10:47amWhen should a team actually switch?
Switch toward open weight when two or more of these are true:
- You are running more than roughly 50 million tokens a month on a task where quality is roughly equivalent (use the SWE-Bench or LiveBench numbers for your specific domain)
- Your data governance or compliance posture requires weights on-premise or in a VPC you control
- You need to fine-tune - either for a proprietary format, a narrow task, or to strip unwanted behaviors
- You have an ML engineer (or a managed inference provider relationship) who can absorb the ops work
Stay with closed APIs when:
- Your workload involves native computer use or rich multimodal inputs
- Traffic is too low to hit the self-hosting break-even utilization threshold
- You need guaranteed SLAs and you are not staffed for infrastructure incidents
- You are doing a prototype - closed APIs ship the fastest path to signal
A practical on-ramp: start with Qwen3-14B locally for evaluation, move to Qwen3-30B-A3B if you need 70B-class performance on constrained hardware, and use a hosted API for production workloads where throughput and latency matter more than cost minimization. For teams that have been locked into closed-source APIs because open-weight alternatives were not competitive enough, that calculus has shifted.
A teammate like Beagle makes this concrete at the task level - it can route different question types to different model backends without requiring your team to rebuild its Slack interface, so you can run the experiment without re-architecting anything.
Open weight vs closed AI model: common questions
Is open weight the same as open source?
No. Open weight means the trained model parameters are publicly downloadable. Open source typically implies the training code, data, and full methodology are also released. Most "open" models - including Qwen3, Llama 4, and DeepSeek - release weights only, under permissive licenses like Apache 2.0 or MIT, but not the full training pipeline.
How much cheaper is a self-hosted open-weight model really?
At the API layer, open-weight models served by inference providers typically cost $0.40-$1.50 per million tokens for frontier-comparable quality, versus $2.50-$15 per million for closed models. Self-hosting eliminates per-token cost entirely but adds GPU infrastructure and ML engineering overhead. The realistic saving for a production workload is 60-80% on token costs, offset by 10-20% additional ops spend.
Can open-weight models handle agentic coding tasks?
Yes, at competitive levels. Qwen3-Coder-480B-A35B scores 67-69.6% on SWE-Bench Verified, within one point of Claude Sonnet 4. For most engineering team workflows - PR review, bug triage, test generation - that difference is below the noise floor of real-world task variation. Complex multi-file, multi-repo tasks still benefit from the closed frontier.
What is the Qwen3 thinking budget and why does it matter?
The thinking budget is a configurable token cap on Qwen3's chain-of-thought reasoning phase. You set a maximum number of tokens the model is allowed to "think" before answering. It lets you trade accuracy for cost on a per-request basis: set a low budget for simple lookups, a high budget for hard reasoning tasks, all from a single deployed model.
Does the open vs closed choice affect my team's AI compliance posture?
Significantly. Closed frontier APIs process your prompts on vendor infrastructure, which creates data-sharing obligations under GDPR, HIPAA, and many enterprise data agreements. Open-weight models self-hosted in your VPC or on-premise mean no data leaves your boundary. Regulated industries - finance, healthcare, legal - often have no real choice: open weight is the only option that clears the legal bar.