AI Forecast Tracker
← Back to forecasts

Open-Source AI Will Break Big Tech's Grip on Intelligence

Will open-weight models consistently match proprietary frontier performance — making AI effectively free — by end of 2027?

If intelligence becomes free, every AI business model built on charging for it needs to be rethought — including the ones your retirement portfolio is betting on.

Target: Dec 2027(447 days until resolution)
Assessed Probability
80%
Very likely
Based on 2 expert predictions, 14 evidence items
Community Forecast
Cast your vote
Be the first to weigh in below

Your Prediction

Where do you think this lands?

Join others who've weighed in

5%95%
50% — More likely than not
Share on LinkedInShare on X
The most underappreciated story of 2026 isn't a new model — it's that intelligence is becoming free. DeepSeek V3.2 delivers frontier-class performance at $0.14 per million input tokens — 27-35x cheaper than leading proprietary models. Chinese open-source models now account for 30% of all global AI downloads, surpassing the US at 15.7%. The MMLU gap between open and proprietary models closed from 17.5 to 0.3 percentage points in a single year. Meta's Llama 4 Scout offers 10 million token context windows with open weights. A RAND report found Chinese models run at one-sixth to one-quarter the cost of comparable American systems. Traditional benchmarks (MMLU, GSM8K) show 90%+ accuracy across all major models — open and proprietary are now indistinguishable on these tests. The implication is profound: if open-weight models match proprietary on capability while costing 10-35x less, the entire pricing power of OpenAI ($730B), Anthropic ($380B), and Google rests on a narrowing moat. The question is whether harder tasks (FrontierMath, ARC-AGI-2) maintain a meaningful gap — or whether open-source catches up there too.

Scenarios

Current value: MMLU gap: 0.3pp (effectively zero). DeepSeek V3.2 at $0.14/M tokens. FrontierMath gap still large (GPT-5.4 at 47.6%, open-source much lower). ARC-AGI-2: best open at ~40% vs proprietary ~83%.

S-curve position: Mid-curve on easy benchmarks (saturated), early curve on hard benchmarks (still large gap)

Bear Case

Persistent 20%+ gap on hard tasks (proprietary compute advantage at frontier is too large for open-source to match)

Base Case

Parity on most practical tasks, 10-15% gap on frontier reasoning — but the practical gap is irrelevant for 90% of use cases

Bull Case

Full parity by mid-2027 (Chinese investment + Llama 5 + DeepSeek V5 close the gap on hard tasks)

How We'll Know

What we measure
Whether open-weight models match proprietary frontier models on hard benchmarks (FrontierMath, ARC-AGI-2, SWE-bench) while costing 10x+ less
Confirmed if
Top open-weight model scores within 5% of best proprietary model on 3+ hard benchmarks (FrontierMath, ARC-AGI-2, SWE-bench) at <10% of the cost
Refuted if
Proprietary models maintain >15% lead on hard benchmarks through end 2027, or the gap widens
Data sources
  • LMSYS Chatbot Arena rankings
  • FrontierMath leaderboard
  • ARC-AGI-2 results
  • SWE-bench Verified
  • Model pricing databases (artificial analysis)

Evidence Trail

Evidence For

  • Mar 9, 2026

    DeepSeek V3.2: $0.14/M tokens (27-35x cheaper). MMLU gap: 17.5pp to 0.3pp in one year. Chinese open-source: 30% of global AI downloads (US: 15.7%). RAND: Chinese models at 1/6 to 1/4 cost. Llama 4 Scout: 10M token context (open weights). 81% of enterprises use 3+ model families — multi-model reality established.→ Probability: 55%

  • Mar 9, 2026

    Gemini 3.1 Pro at $2/$12 per M tokens (lowest price for frontier reasoning). DeepSeek V4 expected with 1T parameters at 32B active — near-zero marginal cost. Traditional benchmarks (MMLU, GSM8K) now 90%+ for all major models. Open-source models can be run on-premise, avoiding API costs entirely. The economic case for proprietary is eroding for all but the hardest tasks.→ Probability: 60%

  • Apr 10, 2026

    MiniMax M2.5 hits 80.2% SWE-bench Verified — within 0.6 points of Claude Opus 4.6 (80.8%), which is the leading proprietary model. GLM-5 scores 77.8%. This is the first open-weight model to cross the 5%-of-best threshold on one of the three resolution-criteria hard benchmarks (SWE-bench). Cost advantage sustained at 27-35x per DeepSeek V3.2 April 2026 data. ARC-AGI-2 gap still meaningful (GPT-5.4 ~83% vs best open ~52%) and FrontierMath remains proprietary-favored — two of three benchmarks still wide.→ Probability: 68%

  • Apr 25, 2026

    DeepSeek V4 preview launched April 24: V4-Pro is a 1.6T-parameter MoE (49B active) — the largest open-weight model to date — with 1M-token context. It hits 80.6% SWE-bench Verified (matching Claude), 93.5% LiveCodeBench (ahead of Gemini 91.7% and Claude 88.8%), 96.4% AIME 2026 (vs Claude 4.6 95.0%), and Codeforces 3206 (ahead of GPT-5.4's 3168). Pricing: $1.74/$3.48 per million input/output — about 5x cheaper than Claude Opus 4.6 and 8-10x cheaper than GPT-5.5. Critically, V4 was confirmed running on domestic Huawei Ascend and Cambricon chips, while V3 had required Nvidia hardware — the export-control workaround is operational, not theoretical. The Stanford HAI 2026 AI Index (April 13) measured the US-China model performance gap at 2.7%, with Chinese models trading top rankings since early 2025.→ Probability: 73%

  • Jun 21, 2026

    GLM-5.2 (Z.ai, June 13) posted 62.1% on SWE-bench Pro and 81.0% on Terminal-Bench 2.1 on third-party leaderboards — the strongest independently verified coding score from any Chinese open-weight lab yet — released within 24 hours of Moonshot's Kimi K2.7 Code (~1T params, open weights). The same week DeepSeek closed its first external round ($7.4B, June 16). Open-weight coding capability is closing on the proprietary frontier from the independently-measured side, not just on saturated benchmarks. Caveat: Kimi's headline scores are self-reported, and the hard-reasoning gap (ARC-AGI-2, FrontierMath) is still wide.→ Probability: 75%

  • Jul 5, 2026

    W27: Meituan open-sourced LongCat-2.0, a 1.6T-parameter MoE and the first Chinese frontier model pre-trained end-to-end on domestic chips — a ~50,000-card cluster with no Nvidia — and it has been topping OpenRouter usage. In the same window Zhipu's ZCode harness put MIT-licensed GLM-5.2 within ~1 point of Claude Opus 4.8 on FrontierSWE at roughly a fifth of the cost. Open-weight capability leadership and hardware self-sufficiency advanced together in a single week. Caveat: "topping OpenRouter usage" is a usage-share signal, not a hard-benchmark parity claim, and the ARC-AGI-2 / FrontierMath reasoning gap is still open.→ Probability: 76%

  • Jul 18, 2026

    W29: The independent leaderboard, not the self-reported table, is what moved this. Moonshot AI released Kimi K3 on July 16 — a 2.8-trillion-parameter open-weight MoE with a 1M-token context, the largest open-weight model ever built — and LMArena's crowd-sourced Frontend Code Arena ranked it #1, a 17-place jump from Kimi K2.6 and ahead of Claude Fable 5. That is a third-party ranking, not a vendor benchmark. Moonshot's own comparison table puts K3 ahead of Opus 4.8 and GPT-5.5 while still behind Fable 5 and GPT-5.6 Sol overall (self-reported, treat accordingly); full weights follow July 27. In the same week Mira Murati's Thinking Machines Lab shipped Inkling, a 975B open-weight MoE with a controllable thinking-effort dial benchmarking above Nvidia's Nemotron 3 Ultra (self-reported) — three labs shipping competitive open-weight frontier models inside seven days. The market treated it as a repricing event: the Philadelphia Semiconductor Index fell 5.7% on July 17 and past a 20% drawdown from its June record, with roughly $3.3T of global semiconductor market capitalisation erased since June 22 and Marvell, ARM and Intel each down more than 30%. Zhipu priced a $4.01B Hong Kong placement on July 8 explicitly to fund MIT-licensed GLM distribution ahead of revenue.→ Probability: 80%

  • Jul 25, 2026

    DeepSeek-V4-Pro and V4-Flash reached general availability July 20 with a 1M-token context and off-peak input pricing near $0.14/M tokens — about two orders of magnitude under Claude Fable 5 — with early testers reporting performance approaching Opus 4.8 (preliminary, not an independent benchmark run). (https://cybernews.com/ai-news/deepseek-and-alibaba-launch-fresh-assaults-on-frontier-ai/)

  • Aug 8, 2026

    W32: Alibaba shipped Qwen3.8-Max on Aug 3 — a 2.4-trillion-parameter (95B-active) MoE with a 1M-token context, the highest-parameter Qwen-Max-class model yet, with open weights promised "next week" (which would make it the first publicly-released rather than API-only Qwen flagship). Alibaba's own table claims OSWorld-Verified 86.1 (vs Claude Fable 5's 85.0) and PaperBench 93.0 (vs 88.8) — but no independent evaluation exists yet, it trails on SWE-bench Pro (67.7 vs 80.0), and Bloomberg framed the release as a "death zone" for US mid-tier model makers. Separately DeepSeek's MIT-licensed V4-Flash-0731 (284B/13B-active) outscored its own larger flagship across nine self-reported agentic benchmarks with an independent Artificial Analysis Intelligence Index of 52. Hold: the parameter and cost story keeps advancing, but the parity claims are self-reported (Benchmark Theater), the SWE-bench Pro gap is real, and Beijing's export-control consultation still threatens the open half of the thesis.→ Probability: 80%

  • Sep 26, 2026

    W39: On ARC Prize's verified ARC-AGI-2 results, open-weight Dots3-Note Preview scored 76.8% at about $0.08 per task against GPT-6 Astra Max at 95.0% and $1.12, cutting the gap from about 31 points in early August to 18 while meeting the cost clause at roughly 7% of the price (https://arcprize.org/results/dots-studio-dots3-note-prev-max).→ Probability: 80%

Evidence Against

  • Mar 9, 2026

    FrontierMath: GPT-5.4 at 47.6%, open-source far behind. ARC-AGI-2: best systems at 83% but require massive compute. Proprietary models maintain lead on the hardest reasoning tasks. Enterprise buyers prefer proprietary for compliance, support, liability. The 0.3pp MMLU gap is on saturated benchmarks — hard tasks show much larger gaps.

  • Jul 18, 2026

    W29 counter-weight to the same events: topping a crowd-sourced frontend-coding arena is one narrow slice of capability, not the three-benchmark parity this topic resolves on, and on Moonshot's own numbers K3 still trails both Claude Fable 5 and GPT-5.6 Sol overall. The financing tells against durability too — Zhipu is raising $4B against roughly $105M of annual revenue and a ~$700M loss, with its stock up ~1,500% since its January IPO, so the Chinese open-weight push is currently funded by public-market enthusiasm rather than commercial traction. Chinese peer stocks fell hard on the K3 release itself (Z.ai -28%, MiniMax -16%), which is competitive pressure inside the open-weight camp rather than a united front against proprietary labs.→ Probability: 80%

  • Jul 25, 2026

    China's Ministry of Commerce spent July consulting Alibaba, ByteDance and Zhipu on new export controls covering training-data transfer and whether foreign users can keep downloading Chinese open-weight models — a tiered regime that could ban public release of the most capable ones, cutting against the China-open-weights-everything assumption. (https://www.tomshardware.com/tech-industry/artificial-intelligence/china-is-considering-export-controls-on-ai-technologies-including-banning-local-companies-from-using-tsmc-report-claims-restrictions-would-also-advanced-ai-models-training-data-and-overseas-acquisitions)

  • Sep 26, 2026

    W39: Epoch AI's FrontierMath runs point the other way - GPT-6 Astra's 93.7% on Tiers 1-3 widened its lead over the best open model (Kimi K3, 72.2%) from about 17 to 21.5 points (https://epoch.ai/benchmarks), and on Artificial Analysis's index the top open-weight model (Xiaomi MiMo-V2.6-Pro, 46) now trails Claude Opus 5.5 (58) by 12 points, up from 8 (https://artificialanalysis.ai/models/open-source).→ Probability: 80%

How Our View Evolved

  • Sep 26, 202680%→80%

    W39 rotation - HOLD at 0.80. Independent data splits: the ARC-AGI-2 gap narrowed to 18 points while FrontierMath widened to 21.5 and the Artificial Analysis index gap to 12. On every benchmark this topic resolves on, open weights still trail the frontier by a steady six to seven months.

  • Aug 8, 202680%→80%

    W32 rotation — HOLD at 0.80. Alibaba's Qwen3.8-Max (Aug 3, 2.4T params, open weights promised next week) is the highest-parameter open-weight-frontier claim yet, with vendor-table parity on OSWorld/PaperBench against Fable 5 — but no independent eval exists, it trails on SWE-bench Pro (67.7 vs 80.0), and DeepSeek's V4-Flash agentic scores are self-reported. Self-reported parity plus the standing Beijing export-control threat keep this from moving above the already-very-high 0.80.

  • Jul 25, 202680%→80%

    Hold at 0.80 after last week's record move. The new signals split: DeepSeek V4 reached general availability July 20 at roughly 1/50th of frontier pricing (off-peak input near $0.14/M tokens), pushing the cost-parity case further, while Beijing began consulting Alibaba, ByteDance and Zhipu on export controls that could restrict public release of the most capable Chinese open-weight models — a direct threat to the open half of the thesis. Kimi K3's arena lead was already priced in W29, so nothing here clears the bar for a further move.

  • Jul 18, 202676%↑80%

    +0.04 → 0.80, the largest move this topic has taken. Moonshot's Kimi K3 (July 16, 2.8T parameters, 1M context) is the largest open-weight model ever built, and LMArena's crowd-sourced Frontend Code Arena ranked it #1 — ahead of Claude Fable 5 — the first time an open-weight model has topped a genuinely independent frontier coding leaderboard rather than a self-reported table. Thinking Machines' Inkling (975B) landed the same week, making three competitive open-weight frontier releases in seven days. The market priced it as a moat event: SOX fell 5.7% on July 17 into a >20% drawdown, ~$3.3T of semiconductor value erased since June 22. Held short of a larger move because a frontend-coding arena is one slice, K3 still trails Fable 5 and GPT-5.6 Sol on Moonshot's own numbers, and Zhipu's $4B raise against ~$105M revenue shows the sector is financed by froth as much as by traction.

  • Jul 5, 202675%↑76%

    +0.01 → 0.76. Meituan open-sourced LongCat-2.0 (1.6T MoE), the first Chinese frontier model pre-trained end-to-end on domestic chips (~50,000-card cluster, no Nvidia), now topping OpenRouter usage; Zhipu's ZCode harness puts MIT-licensed GLM-5.2 within ~1 point of Claude Opus 4.8 on FrontierSWE at ~a fifth of the cost. Open-weight capability and chip self-sufficiency both advanced in one week. Small move given the already-high level; usage-share is not hard-benchmark parity and the reasoning gap remains open.

  • Jun 21, 202673%↑75%

    GLM-5.2 (June 13) hit 62.1% SWE-bench Pro / 81.0% Terminal-Bench 2.1 on third-party leaderboards — strongest independently verified Chinese open-weight coding score yet — alongside Kimi K2.7 Code and DeepSeek's first external round ($7.4B, June 16). Open-weight coding is converging on the frontier from the independently-measured side. +0.02; the hard-reasoning gap (ARC-AGI-2/FrontierMath) is still wide.

  • Apr 25, 202668%↑73%

    DeepSeek V4 Pro launched April 24 as the largest open-weight model ever (1.6T params, 49B active) at near-frontier coding performance (SWE-bench Verified 80.6%, LiveCodeBench 93.5%, AIME 2026 96.4%) for $1.74/$3.48 per million tokens — and confirmed running on domestic Huawei/Cambricon chips. Stanford HAI 2026 AI Index measured the US-China gap at just 2.7%.

  • Apr 10, 202660%↑68%

    MiniMax M2.5 at 80.2% SWE-bench Verified crosses near-parity with Claude Opus 4.6 (80.8%) — first open-weight model to hit one of three resolution-criteria hard benchmarks within 5%. GLM-5 at 77.8% also competitive. Cost advantage sustained 27-35x. ARC-AGI-2 and FrontierMath gaps persist. +0.08 conservative (one of three benchmarks converged).

  • Mar 9, 2026Initial assessment: 60%

    Baseline — initial published assessment. Based on DeepSeek pricing, MMLU gap closure, Chinese open-source adoption data.

What Experts Say

Penn Wharton Budget Model

Nonpartisan economic research, University of Pennsylvania

Track record: 8/10
“AI will increase US GDP by approximately 1.5% by 2035 and roughly 3% by 2055”
Jun 2025 | academic_research
We assess this claim as 60% more likely than not

Mark Zuckerberg

CEO, Meta

Track record: 7/10
“AI agents will write most of Meta's code in the near future”
Jan 2026 | corporate_statement
We assess this claim as 47% more likely than not

What Could Go Wrong

Proprietary models maintain a meaningful capability lead on hard reasoning tasks that matter for enterprise use cases. Enterprise buyers choose proprietary for compliance and support even when open-source is technically equivalent. The cost gap narrows as proprietary models get cheaper too.

What should we track about this topic?