Two AI gaps are narrowing: American open weights and Chinese frontier models

Two gaps are narrowing at once. American open-weight releases from Thinking Machines and Poolside are finally making the cheap, self-hostable model conversation less one-sided. At the frontier, Chinese labs are putting models such as Kimi K3 and GLM-5.2 much closer to the operational territory occupied by Claude, Gemini, Grok, and GPT. We tested thirteen OpenRouter routes on the same 100-case WebBrain browser-planner payload, then ranked them without treating Sonnet 5—or any other single model—as the judge. The result is encouraging, complicated, and much more useful than a one-number leaderboard: inexpensive Chinese text planners lead first-action consensus, Gemini 3.6 Flash is exceptionally clean, tiny-active Poolside Laguna XS is a real American value result, Kimi K3 nearly matches Sonnet 5's dispatch reliability, and Qwen 3.6 27B is still the model we would put in a single RTX 5090 workstation.

The short version

We sent 1,300 comparable requests—100 per model—through the same WebBrain Chrome planner payload. All thirteen result sets completed without a final API error. Instead of measuring “alignment with Claude Sonnet 5,” we compared every model's normalized first action with the other twelve models on each case.

The main findings:

• DeepSeek V4 Flash 0731 ranks first by exact-action peer consensus, followed by Tencent HY3. Both are exceptionally inexpensive text-only planners on the routes we tested, so this is not a complete browser-agent ranking.
• Claude Sonnet 5 remains the cleanest dispatcher by simpler harness measures: 98 schema-valid actions and 47 ideal tool-name choices, both the best in the group. It also cost $7.22 for this replay because no prompt-cache reads were reported.
• Gemini 3.6 Flash is the operational surprise. It emitted 100 schema-valid calls, ranked fourth in exact peer consensus, and held a 1.96-second median while accepting text, image, video, audio, and files.
• Poolside Laguna XS 2.1 is the American open-weight surprise. The 33B/3B-active text model ranks seventh, responds at a 1.15-second median, and cost $0.073 for the replay. Its argument precision still needs work, but this is no longer a novelty row.
• Kimi K3 remains the strongest evidence that the Chinese frontier gap is narrowing. It produced 97 valid calls and 44 ideal tool choices, close to Sonnet's 98 and 47, while agreeing with Sonnet's tool family on 86 cases. It was slow here, and its tested OpenRouter list price was not cheap.
• MiniMax M3 remains our best-balanced Claude-like value generalist, even though it is not the closest model to Sonnet on this narrow first-action measurement. It combines a 1M context, image and video input, a 23B-active MoE, 2.85-second median latency, and a $0.504 replay cost.
• GPT-5.6 needs a two-turn test, not a dismissal. Luna and especially Terra frequently chose get_accessibility_tree, a cautious observation step that this first-action-only harness penalizes. Our empirical experience with GPT-5.6 outside this harness has been quite strong; simply deleting the tree rows would still introduce selection bias.
• Qwen 3.6 27B remains our single-RTX-5090 choice. It is the only dense 27B model in this group, is realistically quantizable into a 32GB consumer GPU, supports vision and video, and avoids per-token API cost. Its current OpenRouter row had schema roughness, so local deployment still benefits from argument validation.

Those conclusions are deployment-specific. A text-only planner, a multimodal cloud agent, and a private one-GPU agent are different products. One global rank hides more than it reveals.

What the thirteen models actually are

“Open source” is often used loosely in model discussions. We use open-weight for downloadable checkpoints and keep hosted closed models separate. Parameter counts below come from the labs' model cards when disclosed; an em dash means the developer has not published a reliable count for that exact model.

The parameter table explains why “small” can be misleading. Laguna XS activates only 3B parameters from a 33B checkpoint. Inkling Small is a 276B-total MoE despite activating only 12B parameters. DeepSeek activates 13B of 284B, HY3 activates 21B of 295B, and MiniMax M3 activates roughly 23B of 428B. Kimi K3 is in a different infrastructure class at 2.8 trillion total and 104B active. Qwen is the outlier in the other direction: all 27B dense parameters participate, but the full model is compact enough to make one-box local inference practical after quantization.

It also shows why DeepSeek and HY3 cannot simply be declared the two best browser models. Their tested routes received textual page state and tool schemas but cannot inspect a screenshot. Laguna XS is also text-only. GLM-5.2's tested OpenRouter route is text-only, although a released Vision NVFP4 checkpoint changes its broader deployment story. Gemini, Grok, MiniMax, Sonnet, Inkling, Qwen, Kimi, and the GPT-5.6 routes have a wider input surface. This benchmark did not send images, video, audio, or files to any model, so multimodality is product context rather than a scored advantage.

What we ran

This was the full-tier, non-frozen planner suite as it existed at commit 7182c21f: 100 Chrome first-action cases per model, 41 available WebBrain tools, native structured tool calls, and no saved request bodies.

cases per model:       100
models:                13
successful requests:   1,300
API errors:            0
retries:               0
concurrency:            3
max output:             4,096 tokens
Act temperature:        0.15
Ask temperature:        0.30
reasoning override:     none
API surface:            OpenRouter Chat Completions
code checkout tested:   7182c21f

The checkout matters. We ran the original ten models on commit 7182c21f, then fast-forwarded the repository from origin/main. When adding Laguna XS, Grok, and Gemini, we detected that the Chrome prompt/tool source had changed. We therefore reran those three in a detached 7182c21f worktree and use only those -7182 result directories in the thirteen-model comparison. The raw files preserve the run metadata; we do not mix post-pull samples into the ranking.

“Non-frozen” also matters. Earlier WebBrain posts used a May 2026 frozen baseline and often reported agreement with a saved Sonnet reference. This run uses one pinned code checkout rather than that historical frozen snapshot. It is appropriate for comparing these thirteen same-payload runs with each other, not for splicing their percentages into older frozen tables.

Finally, this is a first-action test. It asks whether a model dispatches the right opening tool and arguments. It does not let a model browse for twenty turns, recover from a cautious observation, interpret a screenshot, execute code, or revise a plan after seeing a tool result. Those are not footnotes; they materially affect GPT-5.6, multimodal models, and long-horizon agent specialists.

A consensus rank, not a Sonnet rank

For each case, we compared one model with each of the other twelve models. The primary score is the share of those 1,200 pairwise comparisons in which both models chose the same normalized full action: tool name plus normalized arguments. The secondary score compares only the tool name.

This is leave-one-out consensus. Sonnet contributes one peer vote when another model is scored, just like every other model, and cannot grade itself. No model is privileged as the reference.

Consensus is not ground truth. Twelve models can share the same weak habit, and a genuinely better model can disagree for a good reason. That is why the table retains the harness's deterministic ideal-tool and exact-ideal columns. Still, consensus is a much better answer to “what do the models collectively think?” than silently making Sonnet 5 the constitution.

The agreement signal is not noise. All thirteen models selected the same tool family on 30 of 100 cases. At least ten of thirteen agreed on 73 cases. Only one case had a top tool family supported by five or fewer models, and no case tied for the plurality. The mean plurality was 10.89 models out of thirteen.

Some pairwise tool-family agreements were especially strong:

That is the core frontier result. Kimi K3 and Sonnet 5 independently chose the same first tool on 86 of 100 prompts. Grok and DeepSeek also chose the same tool family 86 times, although their exact normalized actions matched only 47 times. Tool-family consensus and argument precision are different capabilities.

DeepSeek and HY3 lead; Gemini changes the middle

DeepSeek V4 Flash is the cleanest cost-performance surprise. It leads exact-action consensus, emits 90 valid calls, matches 39 ideal tool names, and finishes with a 1.56-second median. The complete 100-call replay cost five cents. Its 39.63-second maximum was a single tail outlier; p95 stayed at 5.24 seconds.

HY3 is the steadier second-place result. It is 3.5 consensus points behind DeepSeek, selects 41 ideal tools, and has no schema-invalid output. At $0.246 for the replay, it remains inexpensive enough to put behind retries or a verifier.

Sonnet rises to third when the three new peer models join the consensus pool. It still owns the strongest simple dispatch row: 98 valid calls and 47 ideal tool choices. Gemini 3.6 Flash lands just behind it in fourth with the only 100/100 call rate in the group, a 1.96-second median, and a 3.71-second p95.

GLM-5.2 moves to fifth by exact consensus but retains the group's best exact-ideal count at 20. It is fast at 1.73 seconds median and has a tight 4.44-second p95. Three calls failed schema validation because the accessibility filter contained a stray quote—visible"—which is a small generation defect with a large production consequence if the caller does not validate arguments.

DeepSeek and HY3 share the important downside: their tested routes are text-only. They can plan over URLs, accessibility trees, extracted page text, and prior tool state, but cannot serve as WebBrain's only model when a canvas app, chart, broken accessibility tree, or screenshot contains the decisive information. Gemini can. GLM's tested route cannot, but its separate Vision NVFP4 checkpoint makes that limitation less structural than it first appears.

Gemini 3.6 Flash and Grok 4.5

Google's model page describes Gemini 3.6 Flash as a production model with a 1,048,576-token input window, text/image/video/audio/PDF input, function calling, structured outputs, thinking, and preview computer use. Our OpenRouter sample validates the dispatch side unusually cleanly: 100 emitted calls, 100 schema-valid calls, 36 ideal tool choices, 14 exact ideals, and no API errors.

Gemini still chose get_accessibility_tree first on 59 cases. On its 41 direct-action cases, it selected the ideal tool 36 times—87.8%. Forty-six of its 59 observation calls matched the peer plurality. That profile looks less like a weak router than a fast model whose default policy prefers refreshing state before acting. A two-turn evaluation may materially improve its apparent quality, just as we expect for GPT-5.6.

Grok 4.5 is different. It ranks tenth on exact consensus but second on tool-name consensus at 79.7%. It usually agrees about what kind of action to take, then differs in the arguments. The route produced 94 valid calls, 36 ideal tool names, 17 exact ideals, a 2.61-second median, and a $2.274 replay cost. Forty-five of its 49 accessibility-tree calls matched the peer plurality. It is a coherent, fast frontier planner, but this first-action argument score does not support placing it near the top.

Kimi K3 is the frontier-gap result

Kimi K3 is enormous: 2.8T total parameters, 104B active, 93 layers, and a 1M context architecture. The route completed 97 of 100 cases with valid tool calls, selected the ideal tool 44 times, and agreed with Sonnet's tool family on 86 cases.

Sonnet still leads on the basic dispatch metrics—98 valid calls and 47 ideal tools—and was more than three seconds faster at the median in this sample. Kimi's 7.54-second median and 26.15-second p95 were the slowest in the group. But the quality gap is no longer categorical. On this task, Kimi looks like a frontier multimodal peer with a worse serving profile, not a separate model tier.

Cost needs careful wording. Kimi's observed replay cost was $1.563 versus Sonnet's $7.222, but that does not mean Kimi had the lower list price. At test time, OpenRouter listed Kimi at $3/$15 per million input/output tokens, while Sonnet 5 had introductory $2/$10 pricing. Kimi's run reported 91% of prompt tokens as cache reads; Sonnet reported none. The workload bill favored Kimi because the route cached repeated tool-schema input, not because its headline token price was lower.

That distinction is exactly why we include both price and observed cost.

MiniMax M3: still the best-balanced Claude-like value generalist

Our subjective product view still favors MiniMax M3 as the most attractive Claude-like value generalist in this set. That is not the same as claiming it won this table.

The measured first-action data does not make M3 the closest model to Sonnet 5. They chose the same tool family on 69 cases and the same normalized action on 49. Kimi, GLM, HY3, and DeepSeek all had higher tool-family agreement with Sonnet in this run.

M3's case is broader:

• 428B total parameters but only about 23B active per token;
• text, image, and video input on the tested route;
• roughly one million tokens of context;
• 2.85-second median latency;
• 89 emitted calls, 86 schema-valid;
• a $0.504 observed replay cost.

Three schema failures prevent an unqualified recommendation: an invalid extract_data enum, object-valued clarify options where strings were required, and an invalid press_keys key. Those are fixable with validation and retrying. More importantly, M3 delivers a much wider capability surface than the text-only leaders while costing about one-third of Kimi's observed replay and returning its median response 2.6 times faster.

So the evidence-based wording is: M3 is not the consensus winner and was not the closest to Sonnet on first actions, but it remains our best overall balance of Claude-like agent behavior, multimodality, latency, context, and hosted cost. Longer trajectories, prose quality, visual work, and recovery behavior require a different benchmark.

GPT-5.6 is being cautious, not simply failing

The raw rank puts GPT-5.6 Luna and Terra at the bottom. That is not how they feel in broader use. In our own empirical, day-to-day work, GPT-5.6 performs very well—often extremely well—once it can inspect state, receive tool results, and continue a trajectory. That observation is anecdotal rather than a benchmark score, but it is important context for interpreting a harness that stops precisely at the model's first request for more state.

The suite's ideal first action is never get_accessibility_tree; it expects the model to act from the state already present in the prompt. GPT-5.6 often asks for a fresh accessibility snapshot anyway. Other models frequently agree: Luna's observation step matched the peer plurality in 43 of its 54 tree calls. Terra's matched in 46 of 80. Gemini matched in 46 of 59, and Grok matched in 45 of 49. Tree-first caution is not unique to OpenAI; this test design systematically rewards models that act immediately.

Would the numbers improve if we stripped those rows? Mechanically, yes: Luna's ideal-tool rate becomes 69.6% on the 46 cases where it chose a direct action, and Terra's becomes 60% on 20. But that is selecting the cases after seeing the model's answer. It rewards low coverage and does not reveal what either model would do after receiving the tree.

The correct follow-up is an observation-tolerant two-turn evaluation:

1. accept get_accessibility_tree as a provisional read step;
2. return a deterministic or real tree result;
3. score the next action, total latency, total tokens, and whether the extra observation changed the decision.

There is also an API mismatch worth disclosing. OpenAI's reasoning guide says reasoning models achieve better intelligence and performance through the Responses API than Chat Completions, and recommends passing reasoning items across function calls. We tested OpenRouter's Chat Completions-compatible openai/gpt-5.6-luna-pro and openai/gpt-5.6-terra-pro routes, used no explicit reasoning-effort override, and stopped at the first action. The low rank is real for this slot; it is not evidence that GPT-5.6 performs poorly in a complete agent or in the direct OpenAI stack. Our empirical experience points in the opposite direction.

Luna is nevertheless interesting. OpenAI positions the Luna tier for efficient high-volume work, and its OpenRouter route cost only $0.228 for all 100 calls. If the two-turn test converts those cautious observations into good actions, Luna could become one of the most economical hosted planners in the set.

Reliability: perfect API delivery is not perfect tool delivery

All thirteen comparable result sets completed every HTTP request without a final API error. That is unusually clean. Tool-level dispatch was less uniform.

“No tool” is not an API failure. It means the model answered in prose or stopped without dispatching a structured action. Depending on the prompt, that can be sensible boundary behavior. A schema-invalid action is different: the intent may be obvious to a person, but a strict agent cannot safely execute it.

Qwen's nine validation failures are the biggest operational warning in the suite. Most are variants of a quoted visible accessibility filter, suggesting one compact adapter-side repair could recover much of the row. Until that repair exists, the agent should validate, normalize only unambiguous values, and retry rather than passing malformed arguments to the browser.

Cost: list price and workload price tell different stories

These were the OpenRouter list prices visible at test time, followed by the actual usage.cost sum in the saved responses. Prices can change, and provider routing or cache policy can change even when a model slug does not.

The thirteen comparable replays cost $17.065 in total. Sonnet alone accounted for 42.3% of that bill. DeepSeek delivered the highest exact-consensus score for 0.7% of Sonnet's observed cost, but that ratio depends heavily on DeepSeek's 96.8% cache-read share and Sonnet's zero.

The practical cost lessons are narrower:

• DeepSeek is astonishingly cheap for repeated-schema text planning on this route.
• Laguna XS is almost as striking: the second-cheapest replay, a 1.15-second median, and a genuinely competitive consensus row.
• HY3's low list price works even without a huge reported cache benefit.
• Luna is a credible budget route if a two-turn observation flow validates its behavior.
• Gemini combines the only 100/100 valid-call result with full input multimodality at a middle-of-table workload cost.
• M3 occupies a useful middle: inexpensive, multimodal, and faster than Kimi.
• Grok is fast and coherent at the tool-family level, but its $2.274 replay did not buy top-tier exact-action consensus here.
• Kimi's observed bill was much lower than Sonnet's, but its list price was higher under Sonnet's temporary introductory pricing.
• Local Qwen has hardware and electricity cost, but no metered API token bill and no provider-side data path.

GLM-5.2 Vision is a real ecosystem win

The OpenRouter z-ai/glm-5.2 route in this test is text-only. But Baseten's GLM-5.2-Vision-NVFP4 checkpoint is more important than a speculative “vision may arrive” footnote. It is a released, queryable vision-language artifact that bolts Kimi-K2.6's MoonViT-3d tower onto the frozen GLM-5.2 text backbone through a newly trained 49.5M-parameter PatchMerger projector. The card says neither the 744B-total/40B-active GLM backbone nor the vision tower was modified.

That is a significant win for GLM's ecosystem. The checkpoint has a standard OpenAI multimodal message example, a 1M maximum context, MIT licensing, and ready SGLang/Baseten deployment files. It demonstrates that a strong text-only reasoning backbone can gain sight without retraining the entire model.

The deployment is still infrastructure-scale: roughly 466GB, Blackwell-only, with 8×B200 for 1M context or 4×B200 for 256K. It is not the OpenRouter route we scored, not a single-RTX-5090 model, and we have not run its vision benchmark. It therefore receives no numerical credit in the table.

DeepSeek V4 Flash and HY3 lose product breadth against GLM in this respect: their tested routes are text-only and we do not have an equivalent released vision checkpoint to point to. The GLM method suggests that a similar frozen-backbone vision graft may be technically possible for them. That is an inference, not a shipped capability. Until a checkpoint and evaluation exist, GLM owns this ecosystem advantage.

American open weights are back in the conversation

Laguna XS 2.1 makes the American open-weight claim much more concrete. Poolside's model card describes a 33B-total, 3B-active MoE with 40 layers, 256 experts plus one shared expert, 262K context, downloadable quantizations, and local operation on a Mac with 36GB of RAM. In our same-payload run it ranks seventh in exact consensus, produces 88 schema-valid calls, responds at a 1.15-second median, and costs $0.073.

That is a remarkably efficient row. It is only a text-to-text coding specialist, chooses the ideal tool 31 times, and reaches just five exact ideal actions. One press_keys argument is schema-invalid. So Laguna XS is not a universal browser model or an argument-precision champion. It is, however, the first compact American open-weight model in this cohort that is simultaneously competitive, fast, locally plausible, and almost free to host.

Inkling Small ranks eleventh in exact consensus, emits 84 tool calls, and is not a score champion. Its significance is a different capability surface: a 276B-total/12B-active open-weight preview with text, image, and audio input, 1.05-second median latency on the tested hosted route, and a $0.243 replay cost. Thinking Machines says it will release the full weights after testing is complete.

It also does not stand alone. Our earlier full-size Inkling test verified image and audio input and found a highly parseable, broad model. Poolside's Laguna S 2.1 test showed an extraordinarily inexpensive 118B-A8B American coding model, while Laguna M.1 fixed much of S's no-tool problem at 225B-A23B.

The older Laguna S and M.1 rows came from the frozen suite and must not be numerically inserted into this consensus table. Laguna XS is different: we reran it on the exact 7182c21f payload used here, so its seventh-place position is directly comparable. Six months ago, the cheap open-weight agent conversation was dominated by Qwen, MiniMax, Tencent, DeepSeek, StepFun, and Z.ai. Thinking Machines and Poolside now give US developers credible downloadable options at several scales and with different modality tradeoffs.

The gap is narrowing. It is not closed. Chinese open-weight models hold the first two positions and six of the top nine, while Kimi supplies the strongest open-weight multimodal frontier row. But Poolside now has a $0.073 American model sitting between MiniMax M3 and Kimi K3 on exact peer consensus. That would have been difficult to imagine in the cheap-agent market only a few model generations ago.

Chinese frontier models are closing a different gap

The second gap is between Chinese labs and the leading closed US frontier APIs. This test supplies three different kinds of evidence:

1. Kimi K3 nearly matches Sonnet's dispatch reliability and agrees with its first tool on 86% of cases while remaining open-weight and multimodal.
2. GLM-5.2 leads every model on exact-ideal count, ranks fifth in peer consensus, and now has a credible released path to vision—even though the OpenRouter route tested here is text-only and has three schema failures.
3. DeepSeek V4 Flash and HY3 turn strong planner agreement into commodity-priced inference, again with the text-only limitation.

That is convergence, not equivalence. Sonnet still has the best ideal tool-name count. Gemini is the only model with 100 valid calls and ranks fourth; Grok is fast and has 79.7% tool-family consensus. GPT-5.6 is under-measured by this one-turn design and performs much better in our empirical multi-step use. Vision, audio, computer-use trajectories, coding recovery, safety boundaries, multilingual prose, and provider reliability can all reverse a deployment choice.

Still, “Chinese models are cheaper but clearly a tier behind” is no longer a useful default. In first-action planning, the leading Chinese open-weight routes are peers with distinct modality and serving tradeoffs. Kimi K3 and GLM-5.2 make the frontier gap visibly narrower; DeepSeek and HY3 make the economics harder for every hosted competitor.

What we would deploy for each job

The last row deserves emphasis. Qwen 3.6 27B is still the best model in this set to run on a consumer RTX 5090 box. This new Qwen result used OpenRouter, not our local GPU, so its latency is not a hardware comparison. The hardware recommendation also draws on our earlier local Qwen 3.6 27B NVFP4 planner run, where the model reached a 1.76-second median with native structured tools.

The model is not perfect. The hosted full-suite row has nine schema-invalid calls, and a 27B dense model needs a 4-bit-class quantization rather than BF16 to leave useful room on a 32GB GPU for KV cache and concurrency. But no 284B–2.8T MoE in this table is a realistic single-card alternative. Qwen gives up some hosted frontier quality in exchange for privacy, offline availability, predictable marginal cost, multimodality, and a deployment an individual can actually own.

Bottom line

This benchmark does not produce one universal winner. It produces a much more interesting map.

DeepSeek V4 Flash and Tencent HY3 lead reference-free first-action consensus, but their tested routes are text-only. Claude Sonnet 5 remains the ideal-tool leader at the highest observed replay cost. Gemini 3.6 Flash is fourth with a perfect 100-call validity row and broad multimodality. GLM-5.2 is fifth, leads exact-ideal actions, and gains a major ecosystem advantage from Baseten's released Vision NVFP4 graft. Kimi K3 is close enough to Sonnet on dispatch behavior to make Chinese frontier convergence concrete, although it is slower and not cheap at list price. MiniMax M3 remains our best-balanced Claude-like multimodal value generalist.

GPT-5.6's cautious tree-first behavior demands a fair two-turn test before anyone draws a broad capability conclusion. This is not diplomatic padding: GPT-5.6 performs very well in our empirical multi-step use, and OpenAI explicitly recommends the Responses API over Chat Completions for improved reasoning-model intelligence and performance. The first-action result identifies a harness-policy mismatch worth testing, not a verdict on the model family.

The geopolitical conclusion should be equally precise. American and Chinese open-weight ecosystems are getting closer because Laguna XS now places seventh as a 33B/3B-active, $0.073 text planner and Inkling supplies multimodal breadth. The Chinese ecosystem still leads this particular planner table and has the stronger open-weight multimodal frontier. At the same time, Chinese labs are closing the frontier-model gap: Kimi K3 approaches Sonnet's action reliability, while GLM, DeepSeek, and Tencent make strong planning dramatically cheaper. Gemini's excellent fourth-place row and Grok's high tool-family consensus show that the closed US field is not standing still.

For a hosted system, choose by modality, latency, cache behavior, and tool reliability—not nationality or one benchmark number. For a consumer workstation, the answer is simpler: Qwen 3.6 27B remains the most convincing model here for a single RTX 5090. It is not the largest or the highest-scoring route. It is the model whose capability, modality, ownership, and hardware requirements meet in a machine people can actually put under a desk.

Raw results

The committed result directories are:

test/llm/results/2026-08-02-gpt56-luna-full_chrome_openai_gpt-5.6-luna-pro
test/llm/results/2026-08-02-minimax-m3-full_chrome_minimax_minimax-m3
test/llm/results/2026-08-02-claude-sonnet5-full_chrome_anthropic_claude-sonnet-5
test/llm/results/2026-08-02-glm52-full_chrome_z-ai_glm-5.2
test/llm/results/2026-08-02-deepseek-v4-flash-full_chrome_deepseek_deepseek-v4-flash-0731
test/llm/results/2026-08-02-inkling-small-full_chrome_thinkingmachines_inkling-small
test/llm/results/2026-08-02-qwen36-27b-full_chrome_qwen_qwen3.6-27b
test/llm/results/2026-08-02-kimi-k3-full_chrome_moonshotai_kimi-k3
test/llm/results/2026-08-02-gpt56-terra-full_chrome_openai_gpt-5.6-terra-pro
test/llm/results/2026-08-02-tencent-hy3-full_chrome_tencent_hy3
test/llm/results/2026-08-02-laguna-xs-full-7182_chrome_poolside_laguna-xs-2.1
test/llm/results/2026-08-02-grok45-full-7182_chrome_x-ai_grok-4.5
test/llm/results/2026-08-02-gemini36-flash-full-7182_chrome_google_gemini-3.6-flash

The derived reference-free ranking, token totals, latency percentiles, and cost sums are saved in:

test/llm/results/2026-08-02-full-suite-consensus.json

No request bodies or API keys are stored in those files.

Tags: #OpenWeights #OpenRouter #DeepSeekV4 #TencentHY3 #GLM52 #KimiK3 #MiniMaxM3 #ClaudeSonnet5 #GPT56 #Gemini36 #Grok45 #Poolside #LagunaXS #Inkling #Qwen36 #RTX5090 #ToolCalling #BrowserAgent #WebBrain