Which model is worth it — 2026

• Rebased to Intelligence Index v4.3 on 12 Sept 2026 — every score below is new. AA launched v4.3 on 7 Sept (Terminal-Bench v4.0 and AutomationBench-AA in, τ³-Banking out; category weights unchanged). All 31 models were re-fetched from their live AA pages (4 parallel workers, every anomaly verified directly against its page); scores run ~4–8 points below v4.2 (Fable 5.1 57→53, Spark 1.3 52→45, Flash 41→35). Prices verified stable across the board — no hikes or cuts this round. Estimate tags now survive only on the Agnes pair (Beta 35, 3.0 Flash 36); QFN/GLMF/GLM-5.2 graduated to measured. New: Agnes 3.0 Flash plotted as the cheapest dot ever ($0.03 — x-axis floor widened $0.045→$0.02); y rescaled 30–60 → 24–56.
• GLM-5.3 re-scores to 45 — and now dominates Kimi too. v4.3 takes the 743B build from 49 to 45 at unchanged $1.40/$4.40 rates — 11 over GLM-5.2 (34) at the same money. It dominates Qwen 3.8 Max (45-tie won on price), Kimi K3, GPT-5.6 Terra, Grok 4.5, Claude Sonnet 5, Gemini 3.6 Flash and GLM-5.2. It falls 3 short of Muse Spark 1.3 max (48) and 2 short of GPT-5.6 Sol (47). Off the frontier: beaten cheaper by Spark max ($0.90 vs $0.78).
• Muse Spark 1.2 re-scores to 40 — buried deeper. v4.3 puts 1.2 at Index 40 vs 1.1's 34 — same money, six more points, so 1.1 stays dominated. But 1.3 max posts 48 at the same $0.78, and Gemini 3.8 Flash beats it (41) at $0.58 — 1.2 wins neither comparison. Off the frontier.
• Kimi K3.1 Pro plots at 55 — top open-weights score once weights ship, but one point short of a rung. Released 23 Oct; AA scores it 55 at $2/$10 ($1.54 blended, 90% cache discount), 1M context, 2 API providers. It buries Kimi K3 (44 at $2.31) by 11 points at a third less money, and beats Gemini 4 Argon (53), but Claude Sonnet 5.5 scores 56 at the same $1.54 blend, so the higher Index keeps the rung and Kimi lands grey. Moonshot promises open weights by 4 Nov. A Highspeed variant ($4/$20) has no separate AA score. See the AA model page.
• Ling 3.1 Flash 1019 and MiniMax-M3.1-Flash — two strong budget builds, both dominated. Ling 3.1 Flash 1019 (InclusionAI, 22 Oct) scores 44 at $0.30/$0.90 ($0.19 blended), 3 over the original Ling 3.1 Flash (41) at identical rates — but MiMo-V2.6-Pro scores 46 for $0.18. MiniMax-M3.1-Flash (19 Oct) scores 43 at $0.15/$0.60 ($0.10 blended, 90% cache discount), 14 over MiniMax M3 (29) — but ties Claude Haiku 5.5 at 43, and Haiku is cheaper ($0.08). See AA's pages for Ling 3.1 Flash 1019 and MiniMax-M3.1-Flash.
• DeepSeek V4.1 Flash 1017 plots at 45 — a 6-point jump at identical rates, still dominated. AA scored the 17 Oct build (max effort) on 18 Oct: 45 at $0.30/$1.20 ($0.18 blended, 98% cache discount), 1M context. That buries the original V4.1 Flash row (39) at the same money, but MiMo-V2.6-Pro scores 46 at the same $0.18 blend, so the higher Index keeps the rung and the 1017 build lands grey. See the AA model page.
• Muse Spark Open plots at 39 — dominated by its own family. Meta released it 15 Oct; AA's xhigh page scores 39 at $1.25/$4.25 ($0.78 blended, 88% cache discount) across 6 API providers, 1M context. Muse Spark 1.3 max scores 48 at the same $0.78, so Open lands grey. See the AA model page.
• Three low-end additions — Beam 1.0, Solar Pro 4, Nemotron 3.5 Super VL — all dominated. Each has a live v4.3.2 Index at or above the chart's floor of 26 and a real per-token price. Beam 1.0 (Reflection, 16 Oct): 30 at $0.70/$2.10 ($0.43 blended), 1M context, 4 providers. Solar Pro 4 (Upstage, released 6 Aug, scored on AA 15 Oct): 27 at $0.30/$1.20 ($0.22 blended), 512K context. AA's August launch post cited 42, before the v4.3 rebase; the live v4.3.2 page says 27. Promotional rates on other hosts are ignored. Nemotron 3.5 Super VL (NVIDIA, 15 Oct): 26 at $0.225/$0.575 ($0.14 blended, provider median), 1M context, exactly on the floor. Claude Haiku 5.5 (43 at $0.08) beats all three on both axes. Nemotron 3.5 Bolt (21) is below the floor and not plotted. See AA's pages for Beam 1.0, Solar Pro 4 and Nemotron 3.5 Super VL.
• Muse Wonder plots at 52 — Meta's first frontier-class model, but dominated. Released 12 Oct; AA's max page scores 52 at $4/$20 ($2.94 blended, 95% cache discount), with 1M context. That is 4 points over Spark 1.3 max (48), but it shares Opus 5.5's exact rates and blend, and Opus scores 58, so Wonder lands grey. Sonnet 5.5 (56 at $1.54) also beats it for about half the money. Muse Wonder Full scores 56 on AA, but Meta serves it only to Muse Code subscribers, with no API, so it has no per-token price and stays an explicit blank table row. See the AA model page, AA's launch analysis and Meta's announcement.
• Muse Spark 1.3 moves to max at 48 — and keeps its rung. The max tier opened to general use 5 Sept at unchanged $1.25/$4.25 rates ($0.78 blended) and scores 48 on v4.3 — 8 over 1.2 (40) at the same money, beating Grok 4.7 (46 at $1.35) cheaper and now outscoring GPT-5.6 Sol (47 at $3.08) cheaper too. It keeps the $0.78 frontier rung (48 > GLM-5.3-Flash's 42). See AA's Muse Spark 1.3 analysis and the max model page.
• GLM-5.3-Flash re-scores to 42 — and retakes its rung. v4.3 takes Z.ai's budget build from 46 to 42 at unchanged $0.15/$0.50 — 2 clear of Qwen 3.8 Flash Next (40) at a cent more. The 46-tie is gone; GLM-5.3-Flash is back on the frontier it used to anchor — until Claude Haiku 5.5 (43 at $0.08, 10 Oct) beat it cheaper, so it is off the frontier again.
• Ox Alpha is listed but unplottable — and would break the chart even with a score. A stealth model that appeared on OpenRouter on 20 Aug 2026 from an undisclosed provider, free for both prompt and completion tokens, with a 1M context and strong developer-reported coding results. Two separate blockers: Artificial Analysis has no entry, so there is no Index; and its price is $0, which has no position on a logarithmic cost axis and would yield infinite points per dollar. More importantly, free is a preview subsidy, not a rate — stealth previews historically either graduate to a named model with real list pricing or vanish. Charting it at $0 would imply a permanent economics that nobody has offered.
• Qwen 3.8 Max (0902) re-scores to 45 — still dominated. The 2 Sept build posts 45 at unchanged $2.00/$6.00 ($1.18 blended). GLM-5.3 ties it 45–45 at cheaper money ($0.90), and Muse Spark 1.3 max beats it (48) cheaper — so despite a real price it never reaches the frontier.
• Two names in the list aren't real model tiers. There is no separate "GLM-5.2 Max" model (the API string is glm-5.2; "max" is a reasoning-effort setting), and neither GPT-5.6 Sol's "max" nor Grok 4.5's "high" is a pricing tier — both cost the same per token at any effort level.
• Effort tiers aren't matched across models. AA's default page for each model uses whatever tier it lists as canonical (max, xhigh, high). Cross-model comparison is imperfect for exactly this reason. Only the Agnes pair still carries AA's estimate tag — everything else is now independently evaluated.
• Prices are AA's blended rate (weighted across cached input, input, and output), taken from each model's AA page. AA's aggregate leaderboard page reports materially different blends — the per-model pages are the ones that reconcile with published input/output rates.
• Claude Sonnet 5's price never rose — and v4.3 re-scores it to 38. The $2/$10 rate was slated to go to $3/$15 on 1 Sept 2026, but AA still lists $2/$10 ($1.54 blended). Don't assume the higher rate until AA's per-model page shows it.
• Vendor benchmarks were unusable for scoring. Sonnet 5's SWE-bench Verified is reported as both 72.7% and 85.2% depending on source; most labs have stopped publishing SWE-bench Verified, AIME, and LiveCodeBench altogether in favour of agentic evals. That's why none of them drive the ranking.
• Grok 4.6 v4.3: 44 at $1.35, off the frontier — Sol passes it. Re-scored from 51 (identical $2/$6 rates, 5 points over Grok 4.5's 39). The old domination inverted: GPT-5.6 Sol (47) now beats Grok at higher money, and Muse Spark 1.3 max (48) beats it cheaper. It does not reach Claude Fable 5 (50) or Claude Opus 5 (51).
• Grok 4.7 plots at 46 — dominated, succeeds 4.6 at identical rates. Released 21 Sept; AA's canonical xhigh page scores 46 at $2/$6 ($1.35 blended) — same rates as 4.6, 2 points higher, so 4.6 is now dominated by its successor too (tie-break: equal price, higher Index wins). Muse Spark 1.3 max (48) still beats it cheaper — off the frontier. See the AA model page and xAI's docs ($2/$6 confirmed).
• MiMo-V2.6-Pro plots at 46 — new frontier rung at $0.18. Released 21 Sept; AA scores 46 at $0.435/$0.87 ($0.18 blended, 99% cache discount) — a 20-point jump over V2.5-Pro (26) at identical rates, so the old row is now dominated by its successor. It slots between GLM-5.3-Flash (42 at $0.10) and Spark 1.3 max (48 at $0.78): nothing cheaper scores higher. Cheapest 46 ever plotted, and the top open-weights score on the Index. See the AA model page and Xiaomi's announcement.
• MiMo-V2.6-Flash now scores 38 and takes a budget frontier rung at $0.06. AA added its Index result 27 Sept: 38 at $0.14/$0.28 ($0.06 blended, 98% cache discount), 1M context. Xiaomi's current overseas API price card confirms $0.14/$0.28 and $0.0028 cached input. It beats Agnes Beta (35) at the same $0.06 and GPT-6 Luna (37) at a higher $0.08, so it joins the frontier just after Agnes 3.0 Flash. See the AA model page and Xiaomi's pricing.
• Ling 3.1 Flash scores 41 at $0.19 blended, but is dominated. AA evaluated the 1 Oct release: $0.30/$0.90 with an 80% cache discount gives $0.06 cached input and $0.19 blended. It meets the chart's score floor and has a real API price, but GLM-5.3 Flash scores 42 at $0.10 and MiMo-V2.6-Pro scores 46 at $0.18, so Ling lands grey. AA lists a 1M context window. See the AA model page.
• Claude Haiku 5.5 plots at 43 — a new frontier rung at $0.08. Released 7 Oct 2026; AA's max page scores 43 at $0.10/$0.50 ($0.08 blended, 90% cache discount), with a 1M context. Anthropic's pricing page confirms $0.10/$0.50 and $0.01 cache reads for prompts up to 100K tokens. It beats MiMo-V2.6-Flash (38 at $0.06) by five points for two cents more, and it beats GPT-6 Luna (37 at the same $0.08) on Index. It displaces Qwen 3.8 Flash Next (40 at $0.09) and GLM-5.3 Flash (42 at $0.10), both now dominated. Two caveats: prompts over 100K tokens bill at $0.50/$2.50, roughly five times this blend, and AA counts 440M output tokens across the Index, so cost per task ($0.21) runs well above what the blended figure suggests. The 43 is the max tier; AA's high, medium and low tiers score 38, 34 and 29. See the AA model page and Anthropic's pricing.
• GPT-6 Sol (Daybreak Blue) stays pending — AA has no Intelligence Index for it. AA added the page on 8 Oct as a trusted-access variant (max effort, $2/$10 list rates) and its Cyber Index article says it leads that index, but the model page shows no Intelligence Index, so there is nothing to plot. The charted GPT-6 Sol row (48 at $1.54) stays the reference. See the AA model page.
• Mistral Large 4 Preview plots at 38 — Mistral's first row on the chart, and dominated. Released 6 Oct 2026; AA scores it 38 at $1.36/$4.18 ($0.79 blended, 90% cache discount), with a 524K context and 1T parameters. It clears the chart's floor and carries a real API price, so the plot rule puts it on — but Muse Spark 1.3 max (48 at $0.78) is 10 points smarter for a cent less, and MiMo-V2.6-Flash matches its 38 for $0.06, so it lands grey. It is the most intelligent model from outside the US and China. See the AA model page and AA's launch analysis.
• GPT-6 Sol plots at 48 — tied with Spark max, rung lost on price. Released 22 Sept; AA's max page scores 48 at $2/$10 ($1.54 blended, 90% cache discount) — dead even with Muse Spark 1.3 max (48), and the $0.78 incumbent keeps the rung. It buries its predecessor outright: GPT-5.6 Sol (47) trails by a point at twice the money ($3.08). Its successor GPT-6.1 Sol now scores 52 at a lower $1.47 blend, dominating it too. See the AA model page.
• GPT-6.1 Sol scores 52 at $1.47, but loses its equal-price rung to Argon. Released 29 Sept; AA's canonical max page scores 52 at $2/$10 ($1.47 blended, 95% cache discount), with $0.10 cached input per OpenAI. It improved on GPT-6 Sol at the same uncached rates: 4 points higher with a slightly cheaper blend. Gemini 4 Argon now ties the blend and scores 53, so the higher Index wins the rung. See the AA model page and OpenAI's announcement.
• Gemini 4 Argon plots at 53 — wins the $1.47 tie with GPT-6.1 Sol. Released 30 Sept; AA's high-effort page scores 53 at $2/$10 ($1.47 blended, 95% cache discount), with 1M context. It adds a frontier rung between Spark 1.3 max (48 at $0.78) and Sonnet 5.5 (56 at $1.54), replacing GPT-6.1 Sol at the same price. Access is currently restricted to trusted cyber defenders/testers; Google says broader API rollout is forthcoming. These are introductory rates: Google says they will rise to $4/$20 after the intro period, whose end date is unspecified. See the AA model page and Google's announcement.
• GPT-6 Luna still scores 37 at $0.08, but MiMo-V2.6-Flash now dominates it. Released 22 Sept alongside Sol; AA scores it at $0.10/$0.50 ($0.08 blended, 90% cache discount). The new Flash scores 38 for $0.06, so Luna loses its former frontier rung. Its predecessor paradoxically scores one point higher (GPT-5.6 Luna 38) at more than twice the blend ($0.17). See the AA model page.
• Six backfilled models — all eligible, all dominated. Each has a live AA Index at or above the chart's floor of 26 and a real per-token price, so the plot rule puts them on — and none reaches the frontier: Qwen3.8 2.4T A95B (40 at $2/$6, $1.18 blended — the exact rates of Qwen 3.8 Max, 5 points below it), Gemini 3.1 Pro Preview (30 at $1.74), Quasar 438B (max) (27 at $0.72, based on GLM-5.2), Apodex 1.1 (26 at $0.38 — its Index is now measured, no longer an estimate), GPT-5.5 Instant (June 2026) (26 at $4.35 — the chart's weakest value at 6.0 pts/$) and Kimi K2.7 Code (26 at $0.72). MiMo-V2.5-Pro's label moved left to make room for Apodex at the same Index.
• Step 5 Preview plots at 44 — dominated, no rung. Released 18 Sept; AA's canonical page scores 44 at $1.00/$2.70 ($0.51 blended, 95% cache discount) — but MiMo-V2.6-Pro is 2 points smarter at a third of the blend (46 at $0.18), so Step 5 lands grey. First StepFun model on the chart. See the AA model page.
• Claude Fable 5.1 loses its rung to the 5.5 models. The 1 Sept release still scores 53 at unchanged $10/$50 rates — Sonnet 5.5 scores 56 at $1.54 blended and Opus 5.5 scores 58 at $2.94, so both are smarter and cheaper. Fable 5.1 still dominates its predecessor outright: Fable 5 (50) trails by 3 points at higher money ($7.17 vs $7.70). GPT-6 Astra ties Fable at 53 but loses on price.
• Claude Opus 5.5 still tops the chart at 58. Released 22 Sept; AA's max-with-fallback page scores 58 at $4/$20 ($2.94 blended, 95% cache discount) — 2 clear of Sonnet 5.5 (56) and 5 clear of Fable 5.1/Astra (53). It dominates Opus 5 (51) outright: 7 points up at lower rates ($4/$20 vs $5/$25). See the AA model page and Anthropic's docs ($4/$20 confirmed).
• Claude Sonnet 5.5 takes a new frontier rung at 56. Released 28 Sept; AA's canonical max-with-fallback page scores 56 at $2/$10 ($1.54 blended, 90% cache discount). It sits between Spark 1.3 max (48 at $0.78) and Opus 5.5 (58 at $2.94), so it adds a rung without dislodging either. At the same blend as GPT-6 Sol and Sonnet 5, its higher Index dominates both (48 and 38 respectively). Caveat: AA's first run used a pre-release build with a structured-output bug; Anthropic says the public build fixes it, and AA plans to rerun. See the AA model page, AA's launch analysis, and Anthropic's docs ($2/$10, $0.20 cache read confirmed).
• Eight models carry the frontier under v4.3.2. Agnes 3.0 Flash ($0.03, 36-est) opens as the cheapest dot ever plotted; MiMo-V2.6-Flash ($0.06, 38) takes the next rung; Claude Haiku 5.5 ($0.08, 43) takes the third rung; MiMo-V2.6-Pro ($0.18, 46) remains the cheapest 46 ever plotted; Muse Spark 1.3 max ($0.78, 48) holds the upper-middle; Gemini 4 Argon ($1.47, 53) takes the tie-broken rung; Claude Sonnet 5.5 ($1.54, 56) follows; then Claude Opus 5.5 ($2.94, 58) closes it. The other fifty are dominated — including Kimi K3.1 Pro (55 < 56 at the same $1.54), Ling 3.1 Flash 1019 (44 < 46 at higher blend), MiniMax-M3.1-Flash (43-tie lost to Haiku 5.5 on price), DeepSeek V4.1 Flash 1017 (45 < 46 at the same $0.18), Muse Spark Open (39 < 48 at the same $0.78), Muse Wonder (52 < 58 at the same $2.94), Qwen 3.8 Flash Next (40 < 43 at higher blend), GLM-5.3 Flash (42 < 43 at higher blend), GPT-6 Luna (37 < 43 at the same price), Mistral Large 4 Preview (38 < 48 at ten times the money), Ling 3.1 Flash (41 < 42 at higher blend), Claude Fable 5.1 (53 < 56), Claude Opus 5 (51 < 56), Claude Sonnet 5 (38 < 56), GPT-6.1 Sol (52-tie lost to Argon), GPT-6 Sol (48 < 52 at higher price), GPT-5.6 Sol (47 < 48), Grok 4.7 (46 < 48), GLM-5.3 (45 < 48), Qwen 3.8 Max (45 < 48), Grok 4.6 (44 < 48), Gemini 3.8 Flash (41 < 42), GPT-6 Astra (53-tie lost on price) and Agnes Beta (35).
• GPT-6 Astra re-scores to 53 — shares the old top score, but is now dominated. v4.3 takes Astra from 55 to 53 at unchanged $10/$50 (6.9 pts/$) — tied with Fable 5.1 (53), but Sonnet 5.5 (56) and Opus 5.5 (58) now outscore both. Fable and Astra also both cost more than Sonnet 5.5 ($7.17/$7.70 vs $1.54). Fable 5 (50 at $7.70, 6.5 pts/$) remains dominated too.
• Gemini 3.7 Flash re-scores to 39 — still dominated by its successor. v4.3 takes 3.7 from 45 to 39 at unchanged $0.58; 3.8 sits 2 points above (41) at the exact price. 3.7 still beats its predecessor 3.6 (34) — now on price too, since Google cut 3.6 to the same $0.75/$3.75 list rates (3.7's better cache discount keeps its blend lower: $0.58 vs $0.63). It dominates Grok 4.5 (39 — same Index, cheaper wins), Claude Sonnet 5, Muse Spark 1.1 and GLM-5.2 — but no longer Terra (42), which now outscores it.
• Gemini 3.8 Flash re-scores to 41 — and falls off. v4.3 takes Google's replacement from 47 to 41 at unchanged $0.58 — still 2 over 3.7 (39) at the exact rates, but GLM-5.3-Flash (42) is now cheaper AND smarter. Google loses its rung.
• v4.3.2 micro-rebase: three rows drop a point — and Hy3 falls off. AA's Index is now v4.3.2 (the same 10 evals as v4.3, scores nudged). Live per-model pages verified 25 Sept: DeepSeek V4.1 Flash 40→39, MiniMax M3 30→29, Hunyuan Hy3 26→25. At 25, Hy3 sits below the chart's floor of 26 — the same rule that kept Ling-3.0-flash-VL off — so it moves out of the scatter into an explicit blank row; its rates are unchanged ($0.14/$0.58, $0.11 blended). No frontier member moved.
• GPT-6 Terra still has nothing to plot (checked 1 Oct). AA's GPT-6 family is Astra/Sol/Luna/6.1 Sol — there is no GPT-6 Terra page. No Index means no plot; the model stays pending.
• Solar Mini 4 scores 24, below the plot floor. AA lists $0.10/$0.40 with a 90% cache discount ($0.07 blended), but 24 is below the chart's minimum Index score of 26, so it appears as a blank table row rather than a plotted point. See the AA model page.
• JT-4.1 Flash 236B A21B has an Index score but no usable price. AA added its score on 29 Sept: Index 34 and context 262K. Its live page shows 0 API providers and $0.00 input/output, with no actual token rate. It is above the chart floor but still unplottable on a log-price axis; kept as an explicit blank table row.
• LongCat-2.5-Preview is pending AA scoring (checked 27 Sept). Meituan released it on 25 Sept; its official docs list a 1M context and temporary discounted API rates of $0.30/$1.20 ($0.006 cached input). Artificial Analysis has no model page or Intelligence Index yet, so it cannot be plotted. The launch discount is temporary, so wait for AA's canonical score and durable pricing before adding a point. See LongCat's release notes and pricing docs.
• Scores move — v4.3.2 is now the version throughout. Do not compare any figure here against the August snapshots (v4.1.1), the 5-Sept snapshot (v4.2) or the 12-Sept v4.3 re-fetch. The original 42 plotted rows reflect the 25 Sept verification sweep; Claude Sonnet 5.5 was added 29 Sept, GPT-6.1 Sol on 30 Sept, Gemini 4 Argon + MiMo-V2.6-Flash on 1 Oct and Ling 3.1 Flash on 3 Oct and Mistral Large 4 Preview on 6 Oct and Claude Haiku 5.5 on 7 Oct and Muse Wonder on 13 Oct and DeepSeek V4.1 Flash 1017, Muse Spark Open, Beam 1.0, Solar Pro 4 and Nemotron 3.5 Super VL on 19 Oct and Kimi K3.1 Pro, Ling 3.1 Flash 1019 and MiniMax-M3.1-Flash on 24 Oct. Muse Wonder Full is an explicit blank row (subscription-only, no per-token price). Solar Mini 4 and JT-4.1 Flash are explicit blank rows for below-floor score and missing price respectively. Rows AA no longer lists (GPT-5.6 Sol/Luna, Gemini 3.6/3.7 Flash, Opus 5, Fable 5, Spark 1.1/1.2, GLM-5.2, the Agnes pair) remain at their last verified values.
• DeepSeek V4.1 Flash re-scores to 39 on v4.3.2 — directly verified. Re-fetched 25 Sept from the live page: 39 at $0.30/$1.20 ($0.18 blended), prices unchanged. At 39 it ties Gemini 3.7 Flash — and beats it on price ($0.18 vs $0.58) — but MiMo-V2.6-Pro matches its $0.18 with a 7-point lead, so the Flash row stays dominated grey. Its 17 Oct successor build (1017) now scores 45 at the same rates. The planned 14 Sept routing of deepseek-v4-pro to V4.1 Flash was cancelled, so no succession needed.
• Watch out for wrong GPT-5.6 pricing online — and Sol is off the frontier again. Several pricing blogs list Luna at $1/$6 and Terra at $2.50/$15 per 1M. OpenAI's own API docs say $0.20/$1.20 (Luna) and $2/$12 (Terra) — a 5x error on Luna. AA's Sol cut to $4/$20 ($3.08) plus its v4.3 score of 47 briefly returned it to the frontier — but Spark max (48 at $0.78) now beats it cheaper. The official rates are used here, and Artificial Analysis agrees with them.
• GPT-5.6 Sol, Terra and Luna are three models, not three price tiers of one. They post different Index scores (47 / 42 / 38) — and all three are now OFF the frontier (Sol 47 < Spark max's 48 at $0.78), while Terra and Luna are beaten by cheaper, smarter models.
• Ties are broken on price — and on score. Astra and Fable 5.1 tie at 53 — the cheaper ($7.17 vs $7.70) keeps the top rung. GLM-5.3 and Qwen 3.8 Max tie at 45 — $0.90 beats $1.18. Grok 4.6 and Kimi tie at 44 — $1.35 beats $2.31. Qwen 3.8 Flash Next and Spark 1.2 tie at 40 — $0.09 beats $0.78, both dominated. Gemini 3.7 Flash and DeepSeek V4.1 Flash tie at 39 — $0.18 beats $0.58. MiMo-V2.6-Flash and Agnes 2.5 Pro Beta tie at $0.06 — score 38 beats 35. MiMo-V2.6-Pro, V4.1 Flash and V2.5-Pro tie at $0.18 — 46 wins the slot. Qwen 3.8 Max and Qwen3.8 2.4T A95B tie at $1.18 — 45 wins it. Apodex 1.1, Kimi K2.7 Code, GPT-5.5 Instant and MiMo-V2.5-Pro tie at 26 — $0.18 is cheapest of the four, and all are dominated anyway. GPT-6 Sol ties Spark 1.3 max at 48 — $1.54 vs $0.78, cheaper wins. Gemini 4 Argon and GPT-6.1 Sol tie at $1.47 — 53 vs 52, so Argon takes the rung. Claude Haiku 5.5 and GPT-6 Luna tie at $0.08 — 43 beats 37. The sorter (a.blend - b.blend || b.idx - a.idx) encodes both rules; without the second term a weaker model can land on the line.
• DeepSeek V4 Pro 0813 re-scores to 36 — and stays. v4.3 takes Pro from 42 to 36 at unchanged $1.32/$3.96 — trailing Spark max (48) by 12 at nearly the same money. The planned 14 Sept migration (endpoint serving V4.1 Flash) was cancelled — DeepSeek now says V4 Pro stays with billing unchanged.
• Read DeepSeek's price table by column — prices verified again 12 Sept. DeepSeek's first column is input on a cache hit, not the standard input rate — misreading it once understated V4 Pro's cost ~120x. AA still lists Flash-class at $0.30–$0.44 in / $1.20–$1.32 out and Pro 0813 at $1.32/$3.96 "based on DeepSeek's API". The blended figures here track AA's current pages; if DeepSeek restores promo rates, the rows move left.
• Fast mode is a separate rate. Opus 5 in fast mode bills at $10/$50 per 1M — the same as Fable 5's standard rate, and not what's charted here.
• Qwen3.8-Flash-Next re-scores to 40 — estimate lifted, rung kept. v4.3 takes it from 46-est to 40 measured at unchanged $0.09 — the $0.09 rung (40 > Agnes 3.0 Flash's 36) until Claude Haiku 5.5 (43 at $0.08) displaced it on 10 Oct. Agnes 2.5 Pro Beta re-scores 40→35 (still estimate) at $0.06 — dominated, and now beaten by its own sibling 3.0 Flash on both axes.
• Agnes 3.0 Flash opens the chart at $0.03. Sapiens' new budget build scored 11 Sept at 36 (AA estimate) for $0.05/$0.15 ($0.03 blended, 1M ctx, 237 t/s) — the cheapest dot ever plotted, forcing the x-axis floor wider ($0.045 → $0.02). It takes the opening rung (36 > nothing cheaper exists) at a staggering 1200 points per dollar. If the independent eval moves the estimate, the bottom of the frontier moves with it.

Sources: Artificial Analysis per-model pages (Intelligence Index v4.3.2 throughout) · arena.ai text leaderboard · vendor pricing docs (platform.claude.com, developers.openai.com, docs.x.ai, docs.z.ai, ai.meta.com, deepmind.google, minimax.io, openrouter.ai) · vendor blogs z.ai/blog/glm-5.3-flash & qwen.ai/blog?qwen3.8-flash-next (pricing, not Index). Data as of 24 October 2026.