2026.09.29 | Substack | LLM | Open-Weight Catch-Up Won't Last

Top labs are going to widen their lead again

A succession of strong open-weight models were released this summer. Zhipu launched GLM-5.2 on June 22, followed by GLM-5.3 on August 14, a release focused solely on scaling post-training. On Zhipu’s own evaluations, GLM-5.3 matched or outperformed Fable 5 and GPT-5.6 Sol on several agentic coding tasks. Moonshot AI launched Kimi K3 on July 17. It scored 57 on the Artificial Analysis Intelligence Index, behind only Fable 5 and GPT-5.6 Sol. DeepSeek’s August 13 update to V4 Pro scored 87.9 on Terminal-Bench 2.1, nearly matching Fable 5’s 88 at roughly one-fiftieth of its cost per task. DeepSeek followed with V4.1 Flash in September. Xiaomi released MiMo-V2.6 with open weights on September 22; its Pro variant now leads the open-weight rankings on the Artificial Analysis Intelligence Index.

Many in the industry now put the gap between open-weight and top proprietary models at just a few months. Some argue that distillation is not the main reason open-weight models have kept pace. We agree it is not the only reason, but it has been one of the key ones. What worries investors is whether open-weight developers can keep matching top proprietary models at a fraction of the R&D cost, and if so, how long the proprietary labs’ lead can last. That worry is feeding into AI and compute stocks, as it did after DeepSeek R1 in January 2025.

We do not think open-weight developers can keep catching up this cheaply, and we expect the recent narrowing of the gap to prove temporary, as it did after R1. Distilling capabilities from top proprietary models is getting harder. The latest models from Anthropic, OpenAI and Google no longer expose their full chain of thought as plaintext, and answers and summaries alone provide weaker training signals. Open-weight developers will need to invest more in their own reinforcement learning, synthetic data generation and training environments. All require compute. Over the next few model generations, we expect the leading open-weight models to fall roughly six to twelve months behind top proprietary models again. The same thing happened in 2025: R1 briefly approached o1, then o3, Claude 4 and GPT-5 pulled ahead.

1. Full reasoning now leaves the API only in encrypted form

Anthropic, OpenAI and Google now take a similar approach to hidden reasoning. Their latest models expose a summary of the chain of thought, or no summary at all. Their APIs can also return encrypted reasoning blocks for reuse in later turns: Anthropic uses a signature field, OpenAI uses encrypted_content, and Google uses thought signatures. Developers pass these blocks back unchanged in subsequent requests. The client cannot decrypt them.

Access has tightened in stages. OpenAI has withheld raw reasoning traces since o1 launched in September 2024. Early Gemini reasoning models still exposed them in early 2025, when researchers at Stanford and other institutions reproduced reasoning behavior by fine-tuning on just 1,000 Gemini traces. In May 2025, Gemini 2.5 and Claude 4 switched to summaries. Opus 4.7, released in April 2026, omits even the summary by default. Fable 5.1, released on September 1, does not expose raw reasoning under any setting. Anthropic’s latest models also refuse requests to reveal their hidden reasoning in the answer. Since September 24, Anthropic has charged for refusals classified as reasoning extraction at the model’s standard rates.

Researchers have nevertheless recovered hidden reasoning by exploiting how APIs handle encrypted blocks. An August paper (arxiv.org/abs/2608.09867)) showed that, among the compatible models tested, these blocks could be replayed across sessions and between models from the same provider. The researchers passed a stronger model’s encrypted reasoning to a weaker model, then prompted the latter to transcribe the underlying text. This exploited cross-model compatibility without breaking the encryption itself. Using Claude Haiku 4.5 to extract reasoning generated by Claude Opus 4.8 cost an estimated US$720 per 10,000 traces. The example below comes from the study.

Providers subsequently patched the vulnerability, and Anthropic introduced two restrictions with Fable 5.1. Its encrypted thinking blocks can be reused only by Fable 5.1 or newer models; older models discard them. The blocks are also bound to the full preceding conversation, so changing that context invalidates the signature. This context check is enabled by default for API accounts created on or after August 31. Encrypted reasoning is still returned, but for the latest protected models, such as Fable 5.1, there is no publicly known way to extract the original hidden reasoning. The remaining route is reconstruction: generating approximate traces from the question, answer and summary. The next section looks at how well that works.

2. Summaries are a poor substitute for full reasoning traces

Subscribed Without plaintext reasoning traces, labs attempting distillation must work with answers, summaries and, for agentic tasks, tool-call logs. A Cornell study (arxiv.org/abs/2603.07267)) measured the difference. With DeepSeek R1 as the teacher and Qwen2.5-7B as the student, training on answers alone or answers plus summaries reduced JEEBench accuracy to 21.6% and 24.0%, respectively, from 28.3% without fine-tuning. Training on full reasoning traces raised accuracy to 43.7%. The researchers then trained a trace-inversion model on open-weight model outputs and used it to reconstruct reasoning from each question, answer and summary. Training the student on those reconstructed traces brought accuracy back up to 36.3%, still 7.4 percentage points below the result with full traces.

Source: Zhang, Morris and Shmatikov, arxiv.org/abs/2603.07267 (May 2026 revision); FUNDA.

Reconstruction worked better in a third-party agentic coding experiment reported by Hackread. Using GLM-5.1 as the teacher and GLM-4.7-Flash as the student on SWE-rebench tasks, the researchers reconstructed trajectories from summaries of short reasoning segments. This recovered 85% of the performance improvement achieved by training on full trajectories. Summarizing each entire reasoning trace before reconstruction recovered only 25% of that improvement.

Our industry checks point the same way. Training directly on summaries rarely matches full reasoning traces, and closing that gap takes extra reconstruction, filtering and training. Distillation now costs materially more than it did in the first half of 2026.

3. Catching up now depends on in-house R&D, and the money raised goes into compute

Distilling top proprietary models is becoming a less reliable way for open-weight developers to keep up. They will need to do more of the research and training themselves, including reinforcement learning, synthetic data generation and the development of training environments. These are compute-intensive activities. The technical reports for DeepSeek V4.1 Flash and Kimi K3 emphasize large-scale reinforcement learning and task synthesis. Zhipu described GLM-5.3 as a release devoted entirely to scaling post-training. The leading open-weight developers, our industry checks show, are racing to build out their own training systems and compute capacity.

Fundraising is central to this competition, with most of the proceeds going toward compute. In July, Zhipu raised about HK$31.4 billion through a placement of new H shares. The stated uses include foundation-model R&D and compute infrastructure. DeepSeek, which had previously taken no external funding, raised about RMB 50 billion in its first round this year for foundation models, AI chips and large-scale compute clusters. Reports in late August pointed to a further raise of tens of billions of yuan at a pre-money valuation of roughly RMB 500 billion. Moonshot AI completed a Series F of more than US$3.5 billion, then launched a pre-IPO round at a US$50 billion pre-money valuation. The pressure extends to inference capacity: less than three days after K3 launched, Moonshot suspended new subscriptions, saying demand was approaching its available compute capacity. As distilling the latest teacher models gets harder, open-weight developers’ pace of progress will depend more on the compute they can secure. We expect differences in frontier R&D investment between top proprietary labs and open-weight developers to show up again in the pace of model improvement.

4. After R1, the gap narrowed and then widened again

On January 27, 2025, a week after R1 launched, Nvidia fell 17%, losing US$589 billion in market value. The Philadelphia Semiconductor Index (SOX) dropped 9.2%. DeepSeek’s disclosure of a roughly US$5.6 million training cost for V3 raised an obvious question: were the hyperscalers spending too much? That figure, however, covered only V3’s final training run at assumed GPU rental rates. It excluded the research and experimentation that preceded it.

R1 approached o1’s performance on AIME and SWE-bench, but still trailed it by 4.2 percentage points on GPQA. The full o1 model had been released the previous month. Three months after R1, o3 arrived with scores 11.8, 11.8 and 19.9 percentage points higher on AIME, GPQA and SWE-bench, respectively. Stock performance reflected much more than competition from open-weight models. The April tariff shock pushed Nvidia below its January lows. The stock recovered after the tariff truce in May and stronger-than-expected earnings, regaining its pre-R1 level by June 9. Hyperscalers began raising capex guidance in July, and Nvidia ended the year roughly 30% above its level before the R1 sell-off.

Source: Interactive Brokers; DeepSeek; OpenAI; FUNDA.

The gap between the leading open-weight models and top proprietary models has followed a similar pattern this year. Rescored using the September version of the Artificial Analysis Intelligence Index, K3 scores 44, versus 57 under the old methodology at launch. That puts it six points behind Fable 5, the highest-scoring proprietary model available when K3 launched. By late September, following the release of Opus 5.5, the gap between the top proprietary model and the top open-weight model, MiMo-V2.6 Pro at 46, had widened to 12 points. GLM-5.3 and DeepSeek V4.1 Flash scored 45 and 39, respectively.

Source: Artificial Analysis Intelligence Index v4.3.2, as of September 27, 2026; FUNDA.

Investment view

Our base case is that the gap widens back to six to twelve months and that this summer’s open-weight challenge to top proprietary models proves a one-off, as R1’s did.

For top proprietary labs, that means a longer lead, as their latest capabilities become harder to replicate cheaply. Their enterprise deployments also depend on agent harnesses, the software that manages tools and task execution, and on workflows tailored to individual customers. These are not captured in API outputs, making them difficult to replicate through distillation alone.

We remain bullish on compute and networking infrastructure. Compute demand depends on many factors, and open-weight models are only one of them. As external distillation becomes less effective, open-weight developers must invest more in training their own models, with compute absorbing most of the capital they raise. Top proprietary labs must keep training the next generation to stay ahead. This will keep driving compute demand. Meta, Alphabet and Amazon spent roughly US$292 billion combined on capex in 2025, about 23% above the guidance they issued early that year, shortly after R1 launched. Their combined 2026 guidance is US$545 billion to US$570 billion.

The main risk is agentic work, now the most contested area. Summaries of short reasoning segments, together with tool-call records, still carry much of the training signal: 85% of the gain in the Hackread test. What would change our view is evidence that reconstruction lets open-weight developers keep pace with successive generations of top proprietary models at a materially lower R&D cost. One experiment does not show that.

Invite your friends and earn rewards

If you enjoy FUNDA, share it with your friends and earn rewards when they subscribe.

Invite Friends