A year ago the best downloadable model scored 22 on Artificial Analysis's Intelligence Index. The leading closed model sat at 35. That 13-point gap felt like a structural fact about the industry. As of this month, the highest-scoring open-weight model was DeepSeek V3 0324 at 22 just one year ago, while today Kimi K2.6 and MiMo V2.5 Pro tie at 54 on the same index - within 6 points of GPT-5.5 at 60. That is not incremental progress. That is the gap closing by more than half in twelve months.
The shift matters because the Intelligence Index is not a single benchmark someone can game. Artificial Analysis's v4.1 index incorporates nine independent evaluations: GDPval-AA v2, τ³-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR. That breadth makes it harder to inflate with narrow fine-tuning. A model that scores well across all nine is genuinely capable across a wide range of work tasks - not just one that memorized a test.
What three open-weight releases actually changed
Three releases in a nine-week window drove the shift: Moonshot AI shipped Kimi K2.6 on April 20, 2026; DeepSeek followed with V4 on April 24; and Z.ai closed the run with GLM-5.2 on June 13. Each one leapfrogged the last on a different axis.
GLM-5.2 takes the top open-weight score on the Artificial Analysis Intelligence Index at 51 and leads on SWE-bench Pro at 62.1 percent. DeepSeek V4 Pro is dramatically cheaper per token and carries a full one-million-token context at no premium.
Kimi K2.6 is a 1-trillion-parameter vision-language model designed to generate code in a plan-write-test-debug loop that can last for days, and it can instantiate hundreds of agents that collaborate on a single task.
All three ship weights under permissive licenses - GLM-5.2 under MIT , DeepSeek V4 under MIT with no user cap and no commercial restrictions.
The cost data is the part teams should print out and put on their walls. DeepSeek V4 Pro costs $1,071 to run the full Artificial Analysis Intelligence Index benchmark suite - more than four times cheaper than Claude Opus 4.7 at $4,811.
The three leading open-weight models offer comparable intelligence to leading proprietary models at between half to one-sixth of the price.
Where closed APIs still have a real edge
None of this means you should swap your production Claude or GPT-5.6 API key for a self-hosted cluster this week. The data has important caveats.
DeepSeek V4 Pro and V4 Flash both have a very high hallucination rate - 94 and 96 percent respectively - meaning when they don't know the answer they nearly always respond anyway. That number should stop any team cold before deploying either model on anything customer-facing or facts-critical without a grounding layer.
On the AA-Omniscience benchmark, which measures knowledge reliability and hallucination, the gap to proprietary models remains large: DeepSeek V4 Pro scores -10, MiMo V2.5 Pro scores +4, and Kimi K2.6 scores +6 - compared to GPT-5.5 at +20 and Claude Opus 4.7 at +26.
The infrastructure burden is also real. GLM-5.2's 2-bit quantized variant alone needs roughly 245 GB of combined memory.
Kimi K2.5 and GLM-5.2 require multi-GPU setups with A100 or H100-class GPUs. Teams without a GPU cluster, a dedicated DevOps owner for inference infrastructure, or privacy requirements that justify that investment are still better served by an API.
OpenAI's answer to this pressure arrived July 9. OpenAI unveiled GPT-5.6 in three variants: Sol (the workhorse), Terra (intermediate), and Luna (budget-friendly).
CEO Sam Altman told CNBC that Sol is 54 percent more token-efficient on coding tasks than previous versions.
Pricing lands at $5/$30 per million tokens for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. Luna at $1 input / $6 output is a direct response to the open-weight cost pressure - closed labs are now competing on the open-weight labs' home turf.
Beagle in action#engineering, 10:22amHow to actually pick between self-hosted and closed API
The decision comes down to four things, in order of importance.
Does your data need to stay on your own infrastructure? If yes, open weights are not optional - they are the only answer. A U.S.-headquartered provider offering EU data residency keeps data stored in Frankfurt but governed by U.S. law; an open-weight model deployed on-premises means data never leaves the customer's own infrastructure at all. For regulated industries, that distinction is approaching a hard legal forcing function. EU AI Act enforcement powers activate on August 2, 2026.
Is your workload hallucination-sensitive? If you are summarizing internal documents or generating meeting notes where a wrong fact is merely awkward, the open-weight hallucination rates are manageable with a retrieval-augmented grounding layer. If you are generating customer-facing financial or medical content, the AA-Omniscience gap still matters and closed models hold a real edge.
Do you have multi-GPU infrastructure or a team to run it? DeepSeek V4 is for teams with GPU servers, private cloud infrastructure, or production inference needs. For most developers, starting with Qwen3 or a smaller distill locally and graduating to DeepSeek V4 when the infrastructure justifies it is the smarter path.
Is your primary task agentic coding on a large codebase? That is where open-weight models made their most concrete gains. DeepSeek V4 Flash is the first open-weight model that teams immediately dropped into real agentic pipelines as a plausible substitute for an Anthropic- or OpenAI-class frontier model, and the larger V4 Pro variant set the ceiling with a score of 80.6% on SWE-bench Verified.
The one number that defines this moment
Open-weight models now dominate the Pareto frontier for intelligence vs. price: 9 of the 13 models on that frontier are open-weight. That stat reframes the whole conversation. For years, "use a closed API" was the safe default because it was also the quality default. Those two things have decoupled. Closed APIs remain the convenience default. Open weights are now the value default. Teams that conflate the two are leaving cost and control on the table.
The question is no longer whether open-weight models are good enough. For a growing share of real workloads - internal tooling, agentic coding, document processing on sensitive data - they already are. The question is whether your team has the infrastructure and operational maturity to run them. If the answer is yes, the calculus changed this quarter. If the answer is not yet, building toward it is no longer a moonshot. It is a reasonable eighteen-month plan.