The wire
Material changes only, each carrying the date it was observed and the condition the figure holds under. Frontier movement, price moves at or over 5% on the balanced workload blend, advertised context changes, and published corrections. A new model listing is not a wire event, and neither is anything we merely have an opinion about.
44 material changes
In the 7 days to 2026-09-19 (from 2026-09-13), counted from the events below rather than stamped on the page.
What is not here, and why
- No frontier movement is published for 2026-08-25 → 2026-08-26. The ranked board changed by 62% between those two snapshots — 56 slugs entered the ranked set and 44 left, of a 162-slug union — which is over the 20% ceiling this wire treats as a definition change rather than a market move. The frontier is a Pareto set over that board, so it moves when the board does. Price movement is unaffected: a per-model price delta does not depend on who else is ranked.
- 13 corrections in docs/EMPIRICAL-GROUNDING.md carry no date stamp and are therefore not in the wire. An entry with no date cannot be placed on a timeline. The log itself is complete; this feed is the incomplete view of it.
2026-09-19
frontier join
Z.ai: GLM 5.3 joined the frontier
1475.1 on LMArena at $1.40/M on 2026-09-19, against 1475.1 at $2.15/M on 2026-09-15. Model: z-ai/glm-5.3.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
Google: Gemma 3 4B left the frontier
1290.7 on LMArena at $0.0625/M on 2026-09-19, unchanged from 2026-09-15. Its own figures did not move — the board did. A frontier is a Pareto set, so a model leaves it when the rest of the board changes around it. The wire does not name what now beats it: the dominance rule refuses a replacement that costs capability, and the snapshot projection carries no context window or modalities to check that with. Model: google/gemma-3-4b-it.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
Google: Gemma 4 26B A4B left the frontier
1434.5 on LMArena at $0.1425/M on 2026-09-19, unchanged from 2026-09-15. Its own figures did not move — the board did. A frontier is a Pareto set, so a model leaves it when the rest of the board changes around it. The wire does not name what now beats it: the dominance rule refuses a replacement that costs capability, and the snapshot projection carries no context window or modalities to check that with. Model: google/gemma-4-26b-a4b-it.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
Google: Gemma 4 31B left the frontier
1441.7 on LMArena at $0.1525/M on 2026-09-19, unchanged from 2026-09-15. Its own figures did not move — the board did. A frontier is a Pareto set, so a model leaves it when the rest of the board changes around it. The wire does not name what now beats it: the dominance rule refuses a replacement that costs capability, and the snapshot projection carries no context window or modalities to check that with. Model: google/gemma-4-31b-it.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
OpenAI: gpt-oss-120b left the frontier
1365.9 on LMArena at $0.2625/M on 2026-09-19, against 1365.9 at $0.0703/M on 2026-09-15. Model: openai/gpt-oss-120b.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
OpenAI: gpt-oss-20b left the frontier
1287.3 on LMArena at $0.055/M on 2026-09-19, unchanged from 2026-09-15. Its own figures did not move — the board did. A frontier is a Pareto set, so a model leaves it when the rest of the board changes around it. The wire does not name what now beats it: the dominance rule refuses a replacement that costs capability, and the snapshot projection carries no context window or modalities to check that with. Model: openai/gpt-oss-20b.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
frontier leave
Qwen: Qwen3 30B A3B Instruct 2507 left the frontier
1383.6 on LMArena at $0.0844/M on 2026-09-19, unchanged from 2026-09-15. Its own figures did not move — the board did. A frontier is a Pareto set, so a model leaves it when the rest of the board changes around it. The wire does not name what now beats it: the dominance rule refuses a replacement that costs capability, and the snapshot projection carries no context window or modalities to check that with. Model: qwen/qwen3-30b-a3b-instruct-2507.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-15 → 2026-09-19.
price
DeepSeek: DeepSeek V4 Flash 0423 is 55.6% cheaper on the balanced blend
$0.1089/M on 2026-09-15, $0.0483/M on 2026-09-19 — a 55.6% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-flash.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
DeepSeek: DeepSeek V4 Flash 0731 is 33.3% cheaper on the balanced blend
$0.075/M on 2026-09-15, $0.05/M on 2026-09-19 — a 33.3% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-flash-0731.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
DeepSeek: DeepSeek V4 Pro 0423 is 73.6% cheaper on the balanced blend
$2.00/M on 2026-09-15, $0.5279/M on 2026-09-19 — a 73.6% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-pro.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
DeepSeek: DeepSeek V4 Pro 0813 is 12.2% cheaper on the balanced blend
$0.988/M on 2026-09-15, $0.8672/M on 2026-09-19 — a 12.2% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-pro-0813.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
MythoMax 13B is 45.8% dearer on the balanced blend
$0.06/M on 2026-09-15, $0.0875/M on 2026-09-19 — a 45.8% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: gryphe/mythomax-l2-13b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
MiniMax: MiniMax M1 is 11.7% cheaper on the balanced blend
$0.9625/M on 2026-09-15, $0.85/M on 2026-09-19 — a 11.7% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: minimax/minimax-m1.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Mistral: Mistral Small 3.2 24B is 25% dearer on the balanced blend
$0.1062/M on 2026-09-15, $0.1328/M on 2026-09-19 — a 25% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: mistralai/mistral-small-3.2-24b-instruct.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
MoonshotAI: Kimi K2.7 Code is 5.4% cheaper on the balanced blend
$1.41/M on 2026-09-15, $1.33/M on 2026-09-19 — a 5.4% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: moonshotai/kimi-k2.7-code.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
MoonshotAI: Kimi K3 is 35.9% cheaper on the balanced blend
$5.31/M on 2026-09-15, $3.40/M on 2026-09-19 — a 35.9% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: moonshotai/kimi-k3.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
NVIDIA: Nemotron 3.5 Lightning is 6.8% cheaper on the balanced blend
$0.11/M on 2026-09-15, $0.1025/M on 2026-09-19 — a 6.8% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: nvidia/nemotron-3.5-lightning.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
NVIDIA: Nemotron 3 Nano 30B A3B is 20% dearer on the balanced blend
$0.0875/M on 2026-09-15, $0.105/M on 2026-09-19 — a 20% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: nvidia/nemotron-3-nano-30b-a3b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
OpenAI: gpt-oss-120b is 273.4% dearer on the balanced blend
$0.0703/M on 2026-09-15, $0.2625/M on 2026-09-19 — a 273.4% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: openai/gpt-oss-120b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Qwen: Qwen3 VL 30B A3B Instruct is 13.3% cheaper on the balanced blend
$0.2625/M on 2026-09-15, $0.2275/M on 2026-09-19 — a 13.3% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: qwen/qwen3-vl-30b-a3b-instruct.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Tencent: Hy3 is 37.5% cheaper on the balanced blend
$0.231/M on 2026-09-15, $0.1444/M on 2026-09-19 — a 37.5% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: tencent/hy3.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Z.ai: GLM 5.2 is 60.4% cheaper on the balanced blend
$2.15/M on 2026-09-15, $0.8514/M on 2026-09-19 — a 60.4% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: z-ai/glm-5.2.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Z.ai: GLM 5.3 is 35% cheaper on the balanced blend
$2.15/M on 2026-09-15, $1.40/M on 2026-09-19 — a 35% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: z-ai/glm-5.3.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Z.ai: GLM 5.3 Flash is 40% cheaper on the balanced blend
$0.2375/M on 2026-09-15, $0.1425/M on 2026-09-19 — a 40% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: z-ai/glm-5.3-flash.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-15 → 2026-09-19. Threshold: 5% or more, the same one watch-upstream routes to a human.
correction
Correction: the model page still flattened the ladder that §17 un-flattened
Logged 2026-09-19 as §48 of docs/EMPIRICAL-GROUNDING.md. The wire carries the fact that a correction was published; the log carries what was claimed, what is true, and how it was found. Corrections are not summarised here — shortening one is how a load-bearing qualification gets dropped. A vendor changing a price is not us being wrong; this is.
Correction log, docs/EMPIRICAL-GROUNDING.md §48, logged 2026-09-19.
2026-09-15
frontier join
DeepSeek: DeepSeek V4 Flash 0423 joined the frontier
1431.8 on LMArena at $0.1089/M on 2026-09-15, against 1431.9 at $0.1071/M on 2026-09-11. Model: deepseek/deepseek-v4-flash.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-11 → 2026-09-15.
frontier join
Qwen: Qwen3 30B A3B Instruct 2507 joined the frontier
1383.6 on LMArena at $0.0844/M on 2026-09-15, against 1383.5 at $0.1425/M on 2026-09-11. Model: qwen/qwen3-30b-a3b-instruct-2507.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-11 → 2026-09-15.
frontier leave
Tencent: Hy3 left the frontier
1440.6 on LMArena at $0.231/M on 2026-09-15, against 1439.9 at $0.1444/M on 2026-09-11. Model: tencent/hy3.
Default board: standard delivery, not retired, LMArena general axis, balanced-workload blend. Rank snapshots 2026-09-11 → 2026-09-15.
price
DeepSeek: DeepSeek V4 Flash 0731 is 20% cheaper on the balanced blend
$0.0938/M on 2026-09-11, $0.075/M on 2026-09-15 — a 20% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-flash-0731.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
DeepSeek: DeepSeek V4 Pro 0423 is 68.8% dearer on the balanced blend
$1.18/M on 2026-09-11, $2.00/M on 2026-09-15 — a 68.8% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-pro.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
DeepSeek: DeepSeek V4 Pro 0813 is 13.7% dearer on the balanced blend
$0.8692/M on 2026-09-11, $0.988/M on 2026-09-15 — a 13.7% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: deepseek/deepseek-v4-pro-0813.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Google: Gemma 4 26B A4B is 64.7% dearer on the balanced blend
$0.0865/M on 2026-09-11, $0.1425/M on 2026-09-15 — a 64.7% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: google/gemma-4-26b-a4b-it.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Meta: Llama 4 Maverick is 6.3% cheaper on the balanced blend
$0.324/M on 2026-09-11, $0.3037/M on 2026-09-15 — a 6.3% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: meta-llama/llama-4-maverick.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Meta: Muse Glimmer 30B is 27.5% dearer on the balanced blend
$0.5/M on 2026-09-11, $0.6375/M on 2026-09-15 — a 27.5% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: meta/muse-glimmer-30b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
MoonshotAI: Kimi K3 is 13.4% dearer on the balanced blend
$4.68/M on 2026-09-11, $5.31/M on 2026-09-15 — a 13.4% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: moonshotai/kimi-k3.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
NVIDIA: Nemotron 3 Super is 5.3% dearer on the balanced blend
$0.1638/M on 2026-09-11, $0.1725/M on 2026-09-15 — a 5.3% rise in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: nvidia/nemotron-3-super-120b-a12b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
NVIDIA: Nemotron 3 Ultra is 16% cheaper on the balanced blend
$1.25/M on 2026-09-11, $1.05/M on 2026-09-15 — a 16% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: nvidia/nemotron-3-ultra-550b-a55b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Qwen: Qwen3 14B is 62.3% cheaper on the balanced blend
$0.3981/M on 2026-09-11, $0.15/M on 2026-09-15 — a 62.3% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: qwen/qwen3-14b.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Qwen: Qwen3 235B A22B Instruct 2507 is 60.2% cheaper on the balanced blend
$0.385/M on 2026-09-11, $0.1531/M on 2026-09-15 — a 60.2% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: qwen/qwen3-235b-a22b-2507.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
price
Qwen: Qwen3 30B A3B Instruct 2507 is 40.8% cheaper on the balanced blend
$0.1425/M on 2026-09-11, $0.0844/M on 2026-09-15 — a 40.8% fall in the blended figure. The blend weights input and output, so the list prices behind it may have moved by different amounts and one of them may not have moved at all. Model: qwen/qwen3-30b-a3b-instruct-2507.
Balanced-workload blend, effective price per million tokens, standard delivery. Rank snapshots 2026-09-11 → 2026-09-15. Threshold: 5% or more, the same one watch-upstream routes to a human.
Showing the newest 40 of 131. The rest are in the feed and the JSON.
Where this comes from
Events are computed from the dated rank snapshots in data/snapshots/, a
diff of the upstream models feed, and the correction log. Frontier movement is only
published when the ranked board itself held still between two snapshots — a frontier is
a Pareto set, so a board that changes underneath it moves the answer on its own, and
that is a definition change rather than a market one.
Scores cited are LMArena Elo, used under CC BY 4.0. No Artificial Analysis figures appear on this page or in the feed: their redistribution terms are unresolved. Unrated is not zero. A model with no independent score is named as unrated, never scored.