Between July 17 and August 17, 2026, nine open-weight models shipped. One of them quietly took second place on the leaderboard everyone watches. DeepSeek's V4 Pro 0813 reached #2 overall on Vals AI's independent SWE-bench Verified board at 96.4 — 0.6 points behind Claude Opus 5 (97.0 on the same Vals harness) — at roughly a twelfth of Opus 5's output price. An open-weight model is now effectively tied with the frontier on the coding benchmark that matters most for agentic work.
But the headline number is not the whole story. The clearest finding of the month is how sharply vendor-reported benchmark numbers diverge from independent runs. Same model, same benchmark, sometimes a 33-point gap. And the second theme, less glamorous but more practical: active-parameter counts — not total parameters — now decide what a model costs and where it can run. Three of the nine models fire 3B parameters or fewer per token.
Prices below are OpenRouter averages, USD per 1 million tokens, unless stated. n/p = not published by the lab. n/d = not disclosed. Benchmark provenance is stated for every figure: independent evaluators (Vals AI, Scale AI SEAL, datacurve, Artificial Analysis) are noted where they exist; everything else is vendor-reported.
The SWE-bench Verified board
Vals AI runs the independent SWE-bench Verified harness the labs all cite but rarely run themselves. As of August 17, only three models sit above 96 on Vals' board: Claude Opus 5 (97.0), DeepSeek V4 Pro 0813 (96.4), and GPT-5.6 Sol (96.2, vendor-reported). Kimi K3 lands at 93.4 on the same harness. The open frontier is no longer chasing the closed one on this benchmark — it has caught up.
The benchmark divergence problem
If you only read the labs' own announcements, you would draw the wrong map. The clearest case is DeepSeek V4 Pro on Terminal-Bench 2.1: 87.9 on DeepSeek's own harness vs 54.7 on Vals AI across three trials — the same model, the same benchmark, a 33-point gap. Kimi K3 shows the reverse: 76.8 on Moonshot's SWE-bench run vs 93.4 on Vals AI, the independent number higher by 16.6 points. One lab undersells, one oversells, and you cannot tell which from a press release.
The practical rule this produces: treat a single vendor number as a hypothesis, not a result. When an independent re-run exists, it moves the number — sometimes by a little, sometimes by a third of the scale. Where no independent run exists at all (every Ornith figure, every Muse Glimmer figure, most of the Terminal-Bench scores), the number is a lab's claim about itself.
What they cost to run
The output-price spread across this cohort is 417× — from $0.12 per million tokens (Laguna XS 2.1 and DeepSeek V4 Flash) to $50 (Claude Fable 5). The chart below is log-scaled because the range is otherwise unreadable. DeepSeek quotes off-peak pricing by default; its peak rate is double, so V4 Pro's $1.98 becomes $3.96 at peak. Ornith is self-host only with no published hosted rate, so it sits off this axis entirely — its cost is whatever your own GPUs cost.
Active parameters, not total
Total parameter counts are headline bait. Active parameters per token — how much of the model actually fires on a given token — is what decides cost and where a model can run. The chart below plots both on a log scale because totals span 30B to 2.8T. The two dense models (Qwen 3.8:27B and Muse Glimmer) have active == total, so their bars overlap exactly. Three MoE models — Nemotron Lightning, Laguna XS, and Laguna S — fire between 3B and 8B per token while keeping 30B–118B on disk for routing.
The nine releases
1. Kimi K3 — Moonshot AI, July 16, 2026
The largest open-weight model shipped to date: 1.56 TB of weights, with Moonshot recommending 64+ accelerators to run it, so most users rent rather than self-host. A Mixture-of-Experts architecture with 896 routed experts, of which 16 fire plus 2 shared per token — 104B active of a 2.8T total. 1M context, native text + vision, and tool use. Released under the Kimi K3 License with weights on Hugging Face.
Pricing is input $2.80 / cache read $0.30 / output $14.00 per 1M (first-party $3 / $15). The $0.30 cache read is the practical selling point for long agentic coding sessions: the longer your agent runs, the more of its context is a cache hit, and the cheaper each step gets.
Benchmarks: SWE-bench Verified 93.4 (Vals AI) vs 76.8 (Moonshot's own harness); Terminal-Bench 2.1 88.3 (vendor) vs 80.9 (Vals AI); DeepSWE 1.1 67.5; SWE-bench Pro not published. Artificial Analysis Intelligence Index 57 — 4th overall at launch.
Closest closed comparators: Claude Opus 5 (Jul 24, 2026; $5 → $25; SWE-bench Verified 96.0 vendor / 97.0 Vals; SWE-bench Pro 79.2; Terminal-Bench n/p) and GPT-5.6 Sol (Jul 9, 2026; $5 → $30; Verified 96.2; Pro 64.6; Terminal-Bench 2.1 88.8; DeepSWE 72.7). Opus 5 is the SWE-bench leader K3 is chasing; Sol is the model it effectively ties on Terminal-Bench (88.3 vs 88.8). K3 undercuts both on output price by roughly half.
2. DeepSeek V4 Pro 0813 — DeepSeek, August 13, 2026
The GA build of the 1.6T flagship; same architecture as April's preview, plus DSpark speculative decoding and three reasoning-effort levels. MoE with hybrid CSA + HCA attention, 49B active of 1.6T per token. 1M context, text only (no vision), MIT license.
Pricing (off-peak): input $0.66 / cache read $0.0036 / output $1.98 per 1M, with a peak rate of $1.32 → $3.96. That cache-read price — three-tenths of a cent per million — is what makes long agentic runs cheap.
Benchmarks: SWE-bench Verified 96.4 on Vals AI — #2 overall, #1 open weights; Terminal-Bench 2.1 87.9 (DeepSeek's harness) vs 54.7 (Vals AI, three trials) — the 33-point gap that anchors this report; DeepSWE 1.1 62.7; CyberGym 83.3; SWE-bench Pro not published. Artificial Analysis Intelligence Index 53.
Closest closed comparators: Claude Opus 5 (97.0 SWE-bench Verified on Vals; SWE-bench Pro 79.2) and GPT-5.6 Sol (96.2 Verified; Pro 64.6; Terminal-Bench 88.8; DeepSWE 72.7). They are the only other models above 96 on Vals AI's board. V4 Pro matches them on patch generation at ~1/12 of Opus 5's output price, but loses to both on agentic coding (SWE-bench Pro).
3. Qwen-3.8:27B — Alibaba Qwen, August 14, 2026
The strongest “runs on your own hardware” story of the nine. 27.8B dense (not MoE — all 27.8B active per token), plus an included vision encoder, with Gated DeltaNet attention and multi-token prediction. 262K context, text + image + video in, computer use. Apache 2.0 on Hugging Face and ModelScope. The 55.6 GB checkpoint (~17 GB quantized) runs on a single 24 GB GPU.
Pricing: input $0.40 / output $3.00 per 1M, cache read n/p. Free if self-hosted. Reported to beat Claude Opus 4.6 Max on 15 of 19 overlapping tests.
Benchmarks: SWE-bench Pro 61.7 (Scale AI SEAL leaderboard, independent); DeepSWE 1.1 42.2 (datacurve, independent — roughly triple Qwen3.6-27B); Terminal-Bench 2.1 73.0; LiveCodeBench v6 90.3; OSWorld-Verified 84.3; GPQA Diamond 89.2; SWE-bench Verified not published.
Closest closed comparators: Claude Sonnet 5 (Jun 30, 2026; $2 → $10; SWE-bench Pro 63.2; Terminal-Bench 2.1 80.4) and GPT-5.6 Luna (Jul 9, 2026; $0.20 → $1.20; Terminal-Bench 84.3). Sonnet 5 is the closest closed model on SWE-bench Pro — 63.2 against Qwen's 61.7, a 1.5-point gap from a 27B model on one GPU. Luna is the cheapest closed tier, the real alternative to running Qwen locally for free.
4. DeepSeek V4 Flash 0731 — DeepSeek, July 31, 2026
The best cost-per-capability story of the month for agent fleets. Architecture identical to the April preview; every gain came from a redone post-training run. MoE with 256 routed experts plus 1 shared, 6 firing per token — 13B active of 284B. 1M context, text only, tool use. MIT license; GGUF available with a ~156 GB RAM floor.
Pricing: input $0.06 / cache read $0.0028 / output $0.12 per 1M, first-party $0.14 → $0.28.
Benchmarks: Terminal-Bench 2.1 82.7 (vs 61.8 for the preview build on the same weights); DeepSWE 1.1 54.4; NL2Repo 54.2; CyberGym 76.7; SWE-bench Verified and Pro not published. Artificial Analysis Intelligence Index 50 — above the far larger V4 Pro. GDPval-AA Elo moved 1189 → 1559.
Closest closed comparators: Claude Sonnet 5 (Terminal-Bench 80.4; SWE-bench Pro 63.2) and GPT-5.6 Luna (Terminal-Bench 84.3). Flash beats Sonnet 5 on Terminal-Bench (82.7 vs 80.4) at about 1/83 of its output price, and sits 1.6 points behind Luna at a tenth of Luna's output price.
5. Laguna S 2.1 — Poolside, July 21, 2026
Trained start-to-launch in under nine weeks on 4,096 H200s. NVFP4 weights fit a single DGX Spark. MoE with top-10 of 256 routed experts per token (~6.8% of the model fires), 48 layers, sliding-window attention 3:1 with global — 8B active of 118B. 1M context, text only, thinking toggle with thinking interleaved between tool calls. OpenMDW-1.1 license; free 256K tier on OpenRouter; ~75 GB at Q4_K_M.
Pricing: input $0.09 / cache read $0.01 / output $0.18 per 1M.
Benchmarks: SWE-bench Multilingual 78.5 (category win); Terminal-Bench 2.1 70.2 with max thinking (Artificial Analysis) vs 60.4 with thinking off — always state the mode; SWE-bench Pro 59.4; SWE Atlas 46.2 (Scale AI); DeepSWE 1.1 40.4. Max thinking raises DeepSWE completion tokens from 99k to 249k.
Closest closed comparators: Claude Sonnet 5 (SWE-bench Pro 63.2; Terminal-Bench 80.4) and GPT-5.6 Terra (Jul 9, 2026; $2 → $12; Terminal-Bench 2.1 82.5). Sonnet 5 is the nearest closed model on SWE-bench Pro (63.2 vs 59.4); Terra sits at the same $2 input tier. Laguna S is ~20× cheaper on input and ~55× cheaper on output than either.
6. Muse Glimmer — Meta, August 10, 2026
Meta's first fully open, Apache 2.0 release since the Llama line wound down. Pre-trained on Muse Spark outputs (distillation), shipped with a DFlash drafter for speculative decoding, and demoed running on a MacBook Pro. 30B total — 27.9B decoder + 1.8B vision encoder — dense, all parameters active. 128K context, text + vision, agent-first with tool use. ~17 GB at 4-bit, runs on a 64 GB Mac; Meta serves no first-party API.
Pricing: input $0.30 / output $1.20 per 1M, cache read n/p.
Benchmarks: SWE-bench Verified 76.0; SWE-bench Pro 51.2; Terminal-Bench 2.1 51.7 (weak); MCP-Atlas 75.5; DeepSearch QA 74.6; GAIA2 43.3; AIME 2026 94.7; GPQA Diamond 83.5. Artificial Analysis Intelligence Index 35. All Meta-reported.
Closest closed comparators: Claude Sonnet 5 (SWE-bench Pro 63.2; Terminal-Bench 80.4) and GPT-5.6 Luna (Terminal-Bench 84.3). Sonnet 5 sits in the same coding band, 12 points ahead on SWE-bench Pro; Luna matches Glimmer's hosted output price exactly ($1.20), so the argument for Glimmer is running it locally for free, not the API rate.
Important caveat: Qwen 3.8-27B landed four days later at similar size and beat it across the board; the older Qwen3.6-27B already beat it on Terminal-Bench and OSWorld. Glimmer's wins are on agent suites (MCP-Atlas, DeepSearch QA), not terminal work.
7. Nemotron 3.5 Lightning — NVIDIA, August 11, 2026
NVIDIA's thesis: a frontier model is the wrong tool for most of what an agent actually does (tool calls, retries, validation). Built for the execution layer, with SwitchYard shipped to route work to it. 30B total (Artificial Analysis counts 31.6B) in a hybrid architecture of interleaved Mamba-2, MoE, and attention layers — 3B active per token. 1M context, text only, built for tool calls. OpenMDW-1.1 (weights, training data, and recipes all released); free tier on OpenRouter; runs from Jetson to DGX Spark to data centre.
Pricing: input $0.08 / output $0.20 per 1M, cache read n/p.
Benchmarks: PinchBench 85.4; MMLU Pro 81.9; GPQA Diamond 75.4; SWE-bench Verified 51.6; Terminal-Bench 2.1 24.0 (Artificial Analysis) — the weakest coding score of the nine; SWE-bench Pro not published. Artificial Analysis Intelligence Index 24, level with gpt-oss-120b at a quarter of the parameters. NVIDIA claims up to 4× the output speed of similarly sized models.
Closest closed comparators: Claude Sonnet 5 ($2 → $10) and GPT-5.6 Luna ($0.20 → $1.20) — the tiers a router would escalate to. The comparison is about routing economics, not benchmark parity: Lightning is ~50× cheaper than Sonnet 5 on output and ~6× cheaper than Luna. You buy throughput and volume, not capability.
8. Ornith (ornith-1.0-397b) — DeepReinforce, June 25, 2026
Slightly outside the window, but quantizations and third-party hosting kept landing through July, which is why it belongs in the roundup. Four MIT checkpoints post-trained on Gemma 4 and Qwen 3.5: a 397B MoE flagship, a 35B MoE, a 31B dense, and a 9B dense. The distinctive training method: the model learns to write its own agent scaffold first, then solves the task inside it. Active per token not disclosed; 256K context; text only; MIT license. No hosted rate published — self-host only. The 35B GGUF runs on 8 GB VRAM.
Benchmarks (397B, all self-reported): SWE-bench Verified 82.4; SWE-bench Multilingual 78.9; Terminal-Bench 2.1 77.5; SWE-bench Pro 62.2. The 35B MoE: Terminal-Bench 64.2, SWE-bench Verified 75.6. The 9B: Terminal-Bench 43.1, SWE-bench Verified 69.4. No independent replication published as of August 17, 2026.
Closest closed comparators: Claude Opus 4.8 (May 28, 2026; $5 → $25; SWE-bench Verified 88.6; Pro 69.2; Terminal-Bench 2.1 74.6) and GPT-5.6 Sol (Verified 96.2; Pro 64.6; Terminal-Bench 88.8; DeepSWE 72.7). Ornith beats Opus 4.8 on Terminal-Bench (77.5 vs 74.6) and trails it on both SWE-bench rows; Sol sets the ceiling. Ornith's token cost is whatever your own GPUs cost.
9. Laguna XS 2.1 — Poolside, July 2, 2026
33B total in an MoE with 40 layers, sigmoid gating, mixed sliding-window and global attention at 3:1, and KV cache in FP8 — 3B active per token. 256K context, text only, per-request thinking toggle. OpenMDW-1.1; available via Ollama and llama.cpp; runs on a 36 GB Mac.
Pricing: input $0.06 / output $0.12 per 1M, cache read n/p — tied with V4 Flash as the cheapest output in the cohort.
Benchmarks: SWE-bench Verified 70.9 (from 3B active parameters); SWE-bench Multilingual 63.1 (+5.4% over XS.2); SWE-bench Pro 47.6; Terminal-Bench 2.0 37.5 — note this is the older task set and is not comparable to the 2.1 figures elsewhere; Terminal-Bench 2.1 not published; DeepSWE not published. Poolside ran a reward-hack judge post-hoc and reported no significant hacking.
Closest closed comparators: Claude Sonnet 5 (SWE-bench Pro 63.2) and GPT-5.6 Luna (Terminal-Bench 84.3). Poolside benchmarks XS against the small closed tiers, where the price gap is widest — ~83× cheaper than Sonnet 5 on output, ~10× cheaper than Luna, at roughly 75% of Sonnet 5's SWE-bench Pro score.
Terminal-Bench 2.1: where cost-per-capability is clearest
Terminal-Bench 2.1 measures agentic terminal coding — the workload that agent fleets actually run. On this benchmark, DeepSeek V4 Flash's story is unambiguous: 82.7 at $0.12/1M out beats Sonnet 5 (80.4) at 1/83 the price, and trails Luna (84.3) by 1.6 points at a tenth of the price. Nemotron Lightning's 24.0 is the weakest coding score of the nine — by design, since it is built to route to, not to code with. Laguna XS's 37.5 is on the older 2.0 task set and is excluded from this chart.
Cross-cutting comparison
Three ways to slice the cohort. First, the price ladder — output cost per 1M tokens as a multiple of the cheapest model:
| Model | Output $/1M | × cheapest | Type |
|---|---|---|---|
| Laguna XS 2.1 | $0.12 | 1× | Open |
| DeepSeek V4 Flash | $0.12 | 1× | Open |
| Laguna S 2.1 | $0.18 | 1.5× | Open |
| Nemotron 3.5 Lightning | $0.20 | 1.7× | Open |
| Muse Glimmer | $1.20 | 10× | Open |
| GPT-5.6 Luna | $1.20 | 10× | Closed |
| DeepSeek V4 Pro | $1.98 off-peak | 17× | Open |
| Qwen 3.8:27B | $3.00 | 25× | Open |
| Claude Sonnet 5 | $10 | 83× | Closed |
| Kimi K3 | $14.00 | 117× | Open |
| Claude Opus 5 | $25 | 208× | Closed |
| GPT-5.6 Sol | $30 | 250× | Closed |
| Claude Fable 5 | $50 | 417× | Closed |
| Ornith | self-host only | — | Open |
Second, dense vs. MoE. The dense models are Qwen 3.8:27B and Muse Glimmer (plus Ornith's 9B and 31B checkpoints); the MoE models are Kimi K3, both DeepSeeks, both Lagunas, Nemotron (hybrid Mamba-2 + MoE), and the two larger Ornith checkpoints.
Third, modality: text + vision (Kimi K3, Muse Glimmer); text + image + video (Qwen 3.8:27B); text only (V4 Pro, V4 Flash, both Lagunas, Nemotron, Ornith).
The closed-model reference
Every open-weight model above was measured against a closed comparator. Here is the full closed-tier reference table the comparators draw from:
| Model | Released | In → Out $/1M | SWE-bench Verified | SWE-bench Pro | Terminal-Bench 2.1 | DeepSWE 1.1 |
|---|---|---|---|---|---|---|
| Claude Opus 5 | Jul 24, 2026 | $5 → $25 | 96.0 vendor (97.0 Vals) | 79.2 | n/p | n/p |
| Claude Fable 5 | Jun 9, 2026 | $10 → $50 | 95.0 | 80.0 | 87.2 | n/p |
| Claude Sonnet 5 | Jun 30, 2026 | $2 → $10 | n/p | 63.2 | 80.4 | n/p |
| Claude Opus 4.8 | May 28, 2026 | $5 → $25 | 88.6 | 69.2 | 74.6 | n/p |
| GPT-5.6 Sol | Jul 9, 2026 | $5 → $30 | 96.2 | 64.6 | 88.8 | 72.7 |
| GPT-5.6 Terra | Jul 9, 2026 | $2 → $12 | n/p | n/p | 82.5 | n/p |
| GPT-5.6 Luna | Jul 9, 2026 | $0.20 → $1.20 | n/p | n/p | 84.3 | n/p |
What to actually use
Single GPU / laptop
Qwen 3.8:27B — Apache 2.0, multimodal, 61.7 SWE-bench Pro on one 24 GB card, and it runs for free when self-hosted.
Hosted agent fleets
DeepSeek V4 Flash 0731 — 82.7 Terminal-Bench at $0.12/1M out, 1M context, MIT license. The cost-per-capability leader for volume routing.
Hardest tasks, rented
Kimi K3 — 93.4 SWE-bench Verified on Vals AI, with a $0.30 cache read that keeps long agentic sessions cheap.
What to watch next month
- Meta says Muse Spark 1.2 weights are coming — a US frontier model released in the open, which would be the first of its kind.
- DeepSeek's peak/off-peak billing starts showing up in real invoices. Off-peak quotes look spectacular; peak is double. Real spend will land somewhere in between.
- Independent replays of the August benchmark claims, particularly Ornith's self-reported figures (no independent replication as of August 17) and the Terminal-Bench harness discrepancies that produced the V4 Pro 33-point gap.
A slide-deck version of this report — the visual summary with bullet points for each model — is available as a PDF download at the top of this page.