Accord Partners · Model Intelligence

The open-model gap has collapsed

Open-weight models now trail the frontier by months, not years, and match it on many production benchmarks at half the price.

Updated July 21, 2026
93.4%
SWE-bench Verified
Kimi K3 ranks third overall, ahead of Claude Opus 4.8.
#2
Vals AI Index
74.7%, between Claude Fable 5 and GPT-5.6 Sol.
3–4 months
Average open-model lag
Epoch AI, 2026. The gap was approximately one year in 2024.
Chart 01

Head to head on production benchmarks

Scores are shown within each benchmark. Higher is better.

Chart 02

The price of parity

Token prices mislead across models with different verbosity; cost per completed task is the metric that maps to a budget.

Cost per completed SWE-bench task

Vals AI, measured · Updated 2026-07-17

Kimi K3
$0.21
Grok 4.5
$0.54
GPT-5.6 Sol
$1.15
Claude Opus 4.8
$1.92
Claude Fable 5
$2.05
Median seconds per task
GPT-5.6 Sol182s
Grok 4.5200s
Claude Fable 5356s
Claude Opus 4.8567s
Kimi K3619s

K3 is the slowest in the cluster. Token inefficiency is real; it shows up in wall-clock time, not in cost.

Per token, K3 costs Sonnet-tier money. Per completed job, it is 5 to 10x cheaper than the frontier — the token inefficiency is priced in.

Chart 03

The gap is closing

Estimated lag between the best open-weight and frontier proprietary models.

The remaining edge

Where the frontier still leads

Legal Research Bench

Claude Fable 549.5
GPT-5.6 Sol48.1
Kimi K344.2

Deep-research retrieval still favors frontier models.

DeepSWE

GPT-5.6 Sol73.0
Kimi K367.5

The strongest closed model retains a 5.5-point lead.

Humanity’s Last Exam + vision

Claude Fable 5 lead~10 pts
Kimi K3Trailing

Frontier-hard reasoning and multimodal work remain differentiated.

Parity is benchmark-specific. For frontier-hard reasoning and legal research, closed models keep an edge, for now.

Narrative check

What the bear case gets right

A note circulating among investors makes four claims about open models and AI economics. Here is what public data supports, and what it does not.

“Kimi K3 is not cheap. It is a massive multi-trillion-parameter model that is token-inefficient and requires advanced compute.”

Half right

Right on the sticker: at $3/$15 per 1M tokens, K3 ends the era of discount Chinese AI pricing, and it is the slowest model in the frontier cluster (619s per SWE-bench task vs 182s for GPT-5.6 Sol). Wrong on the conclusion: measured cost per completed task is $0.21 vs $1.15 to $2.05 for the frontier — the token inefficiency is already priced in. The compute-demand read (2.8T parameters, heavy inference hardware to self-host) stands on its own.

“China has not caught up. Western labs hold back their best models, so released comparisons flatter the open ecosystem.”

Unresolved — measured trend points the other way

Internal, unreleased models are not benchmarked; the claim cannot be tested with public data. What can be measured: the released-model gap compressed from about 12 months in 2024 to 3-4 months in 2026 (Epoch AI). Buyers deploy released models, and on that comparison the gap is months and shrinking. The note names its own test: if western releases in the next 45-60 days re-open the gap, this claim gains support. Watch that window.

“It is unclear whether frontier labs make money on models today.”

Pending disclosure — margins improving

Public run-rate reporting is contradictory, but reported inference gross margins have swung from negative in 2024 to strongly positive in 2026. The note is right about the catalyst: IPO disclosures will settle it.

“It is unclear whether hyperscalers earn an adequate return on AI capex.”

Open

Outside the evidence base of model benchmarks. The next several quarters of hyperscaler earnings are the honest test. No verdict offered here.

The core case on this page survives the strongest bear note: whatever exists inside frontier labs, the released open-weight cluster already delivers frontier-grade production benchmarks at the lowest measured cost per job in the market.

For operators

Three implications

Route by workload

Model choice is now a routing decision per workload, not a loyalty decision.

Pricing power breaks

Open weights constrain vendor pricing. Open competition sets the $15 output ceiling.

Deploy with sovereignty

On-prem deployment of frontier-cluster capability is now realistic because the weights are public.

Evidence base

Sources

  1. Vals AI · SWE-bench Verified
    Leaderboard scores, cost per task, and latency.
  2. Vals AI · Kimi K3 model page
    Vals AI Index score of 74.70%.
  3. Vals AI · Benchmarks index
    Production benchmark comparisons.
  4. Vals AI · Legal Research Bench
    Legal research performance.
  5. OpenRouter · Kimi K3
    $3 / $15 pricing and one-million-token context.
  6. Epoch AI · Open-closed ECI gap
    Four-month average lag in May 2026.
  7. Epoch AI · Open versus closed weights
    Approximately three-month lag in October 2025.
  8. Epoch AI · Open models report
    Approximately one-year gap in 2024.
  9. Artificial Analysis · Intelligence Index
    Frontier-cluster intelligence scores.
  10. The Decoder · Kimi K3 analysis
    Pricing and per-task cost analysis.