Current LLM pricing from every major provider — GPT-6 Astra & GPT-6.1 Sol, Claude Opus 5.5, Sonnet 5.5 & Haiku 5.5, Gemini 4 Argon & 3.8 Flash, Meta Muse Wonder, DeepSeek V4.1, Qwen, Kimi, GLM, Mistral Large 4, Grok 4.7, MiniMax. Updated every two weeks because prices don't sit still
| Provider↕ | Model↕ | Input $/M↕ | Output $/M↕ | Context↕ | Strengths↕ | Subscription↕ | Free Tier↕ |
|---|---|---|---|---|---|---|---|
| OpenAI | GPT-6 Astra | $10.00 | $50.00 | 1.05M | Current OpenAI flagship (Sep 3-4 2026) — ends the GPT-5 generation; $10/$50 matches Fable 5.1 exactly; 1,050,000-token context, 128K max output, knowledge cutoff Apr 30 2026; cached input $1.00, cache write $12.50; Batch & Flex 50% off, Fast mode 2x; reasoning.effort adds xhigh and max above high; no fine-tuning; first model to trigger OpenAI's 'critical' cyber tier — production version has restricted cyber caps gated behind verification; prompts over 272K input tokens bill at DOUBLE input ($20) with output at $75 for the whole request; went straight to #1 on Terminal-Bench 4.0 (58.18%) and costs less per solved task than any model above 20% accuracy | LATE SEP 2026: the follow-up GPT-6.1 Astra was SCRAPPED before launch — the WSJ and Reuters reported on Sep 28 2026 that internal testing flagged scope-authorisation failures (pushing ahead without user permission, reaching for external tools when unsafe), so OpenAI shipped the cheaper GPT-6.1 Sol instead and Astra remains the top-rung model. GPT-6 Cyber (enterprise cybersecurity) is promised 'in the coming weeks'. | ChatGPT Plus $20 · Go $8 · Pro 100 $100 (5x Plus) · Pro 200 $200 (20x Plus) · Pro 500 $500 (25x Plus, includes Astra Ultrafast) — Pro sign-ups reopened Sep 29 2026, new Pro 200 subs get the lower ~10x-Plus allowance while grandfathered subs keep the old one to Oct 29 2026 | No free API tier |
| OpenAI | GPT-6.1 Sol | $2.00 | $10.00 | 1.05M | NEW Sep 29 2026 (OpenAI DevDay) — the upgrade to GPT-6 Sol and the new value pick: keeps the same $2 in / $10 out card while CACHED INPUT IS HALVED to $0.10/MTok (5% of input; half of GPT-6 Sol's $0.20 and half of Claude Sonnet 5.5's $0.20), cache writes $2.50; OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work at ONE-FIFTH of Astra's token prices (Astra $10/$50); DeepSWE v1.1 75.2% at high effort versus GPT-6 Sol's best 68.8% at max effort, at ~76% lower cost per task; AutomationBench 31.7% at medium (+4.8 pts); AA Intelligence Index 52 — one behind Astra (53) and behind Opus 5.5 (58); API ID gpt-6.1-sol; live on day one in ChatGPT Work, Codex and the API (NOT in plain Chat); GPT-6.1 Sol Ultrafast (up to 8x token generation in Codex, 6x in the API) announced for 'the coming days'; supersedes GPT-6 Sol outright | ChatGPT Plus $20/mo (chat lives in ChatGPT Work and Codex) | Limited on ChatGPT Free |
| OpenAI | GPT-6 Sol (superseded) | $2.00 | $10.00 | 1.05M | SUPERSEDED Sep 29 2026 by GPT-6.1 Sol, which keeps this exact $2/$10 card and halves cached input to $0.10 — keep 6 Sol only for a pinned integration. NEW Sep 22 2026 — mid tier of the GPT-6 family and the new value pick: OpenAI cut the Sol card in half to $2 in / $10 out, landing on exactly the Claude Sonnet 5 rate card while scoring ~10 points above it on the AA index; cached input $0.20, cache write $2.50; prompts over 272K input tokens bill the whole request at $4 in / $0.40 cached / $5 cache write / $15 out; Batch & Flex 50% off, Fast mode 2x; 1,050,000-token context, 128K max output; API ID gpt-6-sol; live in ChatGPT Work and Codex on paid plans (Sol 50 credits/1M in, 250/1M out); built for complex coding and agentic workflows; supersedes GPT-5.6 Sol outright — a permanent rate below even the old promo | ChatGPT Plus $20/mo | Limited on ChatGPT Free |
| OpenAI | GPT-6 Luna | $0.10 | $0.50 | 1.05M | NEW Sep 22 2026 — the NEW PRICE FLOOR of the mainstream market: $0.10 in / $0.50 out, down from GPT-5.6 Luna's $0.20/$1.20 (50% off input, 58% off output); cached input $0.01; shares Sol's 1,050,000-token context and 128K max output, knowledge cutoff May 18 2026; built for high-volume clearly-shaped jobs — summarisation, field extraction, short-answer work (Luna 2.5 credits/1M in, 12.5/1M out in ChatGPT Work); API ID gpt-6-luna; the cheapest row on most September trackers (Claude Haiku 5.5 matched this exact card on Oct 7 2026); OCT 2026: an updated 'GPT-6 Luna (October)' snapshot began rolling out to ChatGPT Free/Go on Oct 8 alongside new October Sol in paid tiers — API rate card unchanged | ChatGPT Plus $20/mo | Limited on ChatGPT Free |
| OpenAI | GPT-5.6 Sol (legacy) | $4.00 | $20.00 | 1.05M | LEGACY since Sep 22 2026 — superseded by GPT-6 Sol at $2/$10, which is permanently half this rate and beats the GPT-5.6 Sol promo ($4/$20 through Nov 21 2026, list $5/$30) described here; cached input $0.40; >272K input bills $8 in / $30 out; Batch halves; still the reasoning leader (GPQA-D 94.6%). Keep only for pinned integrations — new work should move to gpt-6-sol | ChatGPT Plus $20/mo | Limited on ChatGPT Free |
| OpenAI | GPT-5.6 Terra (legacy) | $2.00 | $12.00 | 1.05M | LEGACY-ISH since Sep 22 2026 — no GPT-6 Terra has shipped, but GPT-6 Sol now sits at the same $2 input with a $10 (not $12) output, so Terra is dominated on this board; mid tier after the Jul 30 2026 cut ($2.50/$15 → $2/$12); 1.05M ctx; Terminal-Bench 2.1 87.4%; the $2 input tier is the current price war zone | ChatGPT Plus $20/mo | Limited on ChatGPT Free |
| OpenAI | GPT-5.6 Luna (legacy) | $0.20 | $1.20 | 1.05M | LEGACY since Sep 22 2026 — was the price floor after the Jul 30 2026 80% cut ($1/$6 → $0.20/$1.20), now undercut by GPT-6 Luna at $0.10/$0.50; a budget price on a current-generation family rather than a legacy model, but no longer the cheapest mainstream row | ChatGPT Plus $20/mo | Limited on ChatGPT Free |
| OpenAI | GPT-4.1 | $2.00 | $8.00 | 1M | Legacy workhorse; 1M ctx; partly retired — the Oct 23 2026 shutdown covers gpt-4.1-nano (→ gpt-5.6-luna), gpt-4o-2024-05-13, o1, o1-pro, o3-mini (→ gpt-5.6-sol) and o4-mini (→ gpt-5.6-terra); gpt-4.1 and gpt-4.1-mini themselves have NO announced shutdown date as of Sep 19 2026 | ChatGPT Plus $20/mo | ChatGPT Free limited |
| OpenAI | GPT-4o | $2.50 | $10.00 | 128K | Legacy, price grandfathered; the gpt-4o-2024-05-13 snapshot retires Oct 23 2026 (→ gpt-5.6-sol); gpt-4o-transcribe still $2.50/$10 | ChatGPT Plus $20/mo | ChatGPT Free limited |
| OpenAI | o3 (Reasoning) | $2.00 | $8.00 | 200K | Legacy reasoning tier at post-cut rates; o3-pro $20/$80 above it; cost-sensitive reasoning now goes to GPT-5.6 Sol's thinking mode or Terra | ChatGPT Pro $200/mo | Limited on ChatGPT Free |
| OpenAI | o4-mini | $1.10 | $4.40 | 200K | Legacy cheap reasoning; fast at coding & STEM; DEPRECATED — o4-mini-2025-04-16 shuts down Oct 23 2026 and fine-tuned ft-o4-mini-2025-04-16 migrates to gpt-5.6-terra; superseded on price by GPT-5.6 Terra ($2/$12) and Luna ($0.20/$1.20) | ChatGPT Plus $20/mo | Limited |
| Anthropic | Claude Fable 5.1 | $10.00 | $50.00 | 1M | Current #1 frontier model (Sep 1 2026); CACHE READS CUT 75%: $1.00 → $0.25/MTok (0.025x input, vs 0.1x industry norm) — ~25% cheaper typical, up to ~45% agentic; 5m cache write $12.50, 1h $20; no long-context surcharge on the full 1M window; Batch $5/$25; AA Intelligence Index 66; gated twin Mythos 5.1 (same weights, $10/$50, cyber defenders only) | Claude Pro $20/mo | Claude Free limited |
| Anthropic | Claude Opus 5.5 | $4.00 | $20.00 | 1M | NEW Sep 22 2026 — Anthropic's first 5.5-family model and the new front-runner: every rate on the card is below Opus 5, with input/output down 20% ($5/$25 → $4/$20) and CACHE READS DOWN 60% ($0.50 → $0.20), which Anthropic frames as ~40% lower cost on typical workloads once both effects land; fast mode $8/$40, Batch $2/$10; full 1M-token window billed at standard rates with no long-context surcharge; thinking CANNOT be disabled; available via API, Bedrock, Vertex AI and Microsoft Foundry, and on Claude Pro / Max / Team / Enterprise (not Free); Anthropic says it performs at Fable 5.1's level on most work and is >30% faster than Opus 5, at 60% below Fable 5.1 on input and output; Sonnet 5.5 shipped Sep 28 2026 and Haiku 5.5 Oct 7 2026 — the 5.5 family is now complete | Claude Pro $20/mo | Not on Claude Free |
| Anthropic | Claude Sonnet 5.5 | $2.00 | $10.00 | 1M | NEW Sep 28 2026 — second model in the Claude 5.5 family, six days after Opus 5.5, and the same $2/$10 card as Sonnet 5. CACHE READS HALVED Oct 7 2026 (announced with Haiku 5.5): $0.20 → $0.10/MTok — Anthropic says that makes most agentic work ~20% cheaper on Sonnet 5.5; 5-min cache writes $2.50, 1-hour $4.00, Web Search $10/1K calls, Batch halves to $1/$5; 1M-token context with no long-context surcharge, 128K max output (300K beta on Batch), adaptive thinking with default effort high, knowledge cutoff Jun 2026; API ID claude-sonnet-5-5 (Bedrock anthropic.claude-sonnet-5-5), live on the Claude API, Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry; Anthropic claims output 30%+ faster and up to 30% lower cost per task than Sonnet 5 because it needs fewer tokens (Artificial Analysis measures 137 tok/s at Max against 75); AA Intelligence Index 56 at max — the exact rate card OpenAI matched with GPT-6.1 Sol the next day; Haiku 5.5 shipped Oct 7 2026 and completes the 5.5 family | Claude Pro $20/mo | Claude Free limited |
| Anthropic | Claude Opus 5 (legacy) | $5.00 | $25.00 | 1M | SUPERSEDED Sep 22 2026 by Opus 5.5 at $4/$20 — 20% cheaper on both ends with cache reads down 60%; still listed here because it was the value pick for hard reasoning (3x cheaper than Opus 4's $15/$75); cache hits $0.50; fast mode $10/$50; matches the Opus 4.8 rate | Claude Pro $20/mo | Limited |
| Anthropic | Claude Sonnet 5 (superseded) | $2.00 | $10.00 | 1M | $2/$10 made PERMANENT on Aug 11 2026, and SUPERSEDED Sep 28 2026 by Claude Sonnet 5.5 at the same $2/$10 card (see that row) — pin Sonnet 5 only for a frozen integration. The scheduled Sep 1 step-up to $3/$15 was cancelled (first headline increase to evaporate before taking effect); cache read $0.20; the coding/agentic default and the rate card OpenAI landed exactly on with GPT-6 Sol on Sep 22 2026 and again with GPT-6.1 Sol on Sep 29. | Claude Pro $20/mo | Claude Free limited |
| Anthropic | Claude Haiku 5.5 | $0.10 | $0.50 | 1M | NEW Oct 7 2026 — completes the Claude 5.5 family (Opus 5.5 Sep 22, Sonnet 5.5 Sep 28, Haiku 5.5 Oct 7) and Anthropic's cheapest/fastest small model ever: $0.10 in / $0.50 out per MTok for prompts up to 100K tokens — 90% below Haiku 4.5 and the SAME card as GPT-6 Luna; above 100K input the whole request bills $0.50 in / $2.50 out (still 50% below 4.5); cache reads $0.01/$0.05, cache writes $0.125/$0.625; 1M-token context (up from 200K on 4.5) with up to 128K output (vs 64K); ~75% lower average cost than Haiku 4.5 per Anthropic (updated tokenizer means slightly more tokens per task, which the average accounts for); first Haiku with adjustable effort (low→max); Terminal-Bench 4.0 39.2% (vs Haiku 4.5's 0.0%), OSWorld 2.1 72.4% offline subset; API ID claude-haiku-5-5, live on all platforms (AWS, Google Cloud, Azure); Haiku 4.5 ($1/$5, cache read $0.10) remains for pinned integrations but is now 10x the price at this tier | Claude Pro $20/mo | Claude Free limited |
| Gemini 3.8 Flash | $0.75 | $3.75 | 1M | NEWEST Flash (Sep 2 2026) — 4th Flash in under 4 months; Terminal-Bench 2.1 jumped 81.6% → 90.8%; intro rate $0.75/$3.75 (thinking included) through Dec 31 2026 then DOUBLES to $1.50/$7.50 on Jan 1 2027; cache read $0.075 (doubles to $0.15 in 2027) with cache STORAGE $0.50/1M/hr through 2026; Batch/Flex 50% off; Priority $1.35/$6.75; gated twin Gemini 3.8 Flash Cyber (vulnerability hunting) is Fairwind Program only — a rate card for it now appears on Google's Agent Platform pricing page, but access is still invitation-only; Sep 15 2026 added the Gemini 3.8 Live / Live Extended Thinking pair (speech-to-speech, no published per-token rate); works harder per task (~40% higher cost/task than 3.7 at same token price) | Google AI Pro $19.99/mo | Gemini Free limited | |
| Gemini 4 Argon | $2.00 (intro to Nov 30) | $10.00 (intro to Nov 30) | 1.05M | GENERALLY AVAILABLE Oct 13 2026 (announced Sep 30 to Fairwind cyber defenders only) — Google's first Gemini 4 model and the current frontier flagship, now purchasable: model ID gemini-4-argon, live on the paid Gemini API, AI Studio and the Gemini platform, on Antigravity for AI Ultra, and in the Gemini app plus AI Mode in Search for AI Ultra AND Pro subscribers; OpenRouter listed google/gemini-4-argon on Oct 13 at the same $2/$10. PRICE IS TEMPORARY: $2 in / $10 out / $0.10 cached through Nov 30 2026, then $4/$20/$0.20 from Dec 1 2026 (Google's own pricing page, updated Oct 16 — verify before budgeting, the standard rate equals what Meta charges for Muse Wonder). Unified 1,048,576-token window covering input + thinking + function calls + output; long-decode output ceiling up to 1,000,000 tokens minus input used (previous Gemini cap 64K), default sync 65K; thinking levels low/medium/high; Batch/Flex half ($1/$5), Priority $3.60/$18; implicit caching only (no explicit cache); grounding $14 per 1,000 requests after 5,000 free/month shared across Gemini 3+. Benchmarks: AA Intelligence Index 52.6 at high effort (the 53 quoted at launch was the pre-restatement index), #1 on AutomationBench-AA at 77.5%, 15% hallucination on AA-Omniscience — lowest of any model above 45 — at ~$1.99 per index task against Astra's $3.26, though it burns ~62K output tokens per task to Astra's 27K; vendor-reported DeepSWE v1.1 77.9%, LVBench 91.7%, CWE-bench 68%. No free tier. | Included for Google AI Pro $19.99/mo and AI Ultra ($99.99/$199.99) in the Gemini app and AI Mode in Search since Oct 13; Antigravity access for AI Ultra | No free tier (paid API only) | |
| Gemini 3.7 Flash | $0.75 | $3.75 | 1M | Launched Aug 13 2026; same intro rate as 3.8 through Dec 31 2026 ($1.50/$7.50 from Jan 1 2027); still supported — the cheaper-per-task fallback when 3.8's token hunger bites; cache $0.075 with $0.50/1M/hr storage; Batch/Flex $0.375/$1.875; Priority $1.35/$6.75 | Google AI Pro $19.99/mo | Gemini Free limited | |
| Gemini 3.1 Pro | $2.00 | $12.00 | 1M | UNRESOLVED NAMING CONFLICT (re-checked Oct 3 2026): Google Cloud's own model page for Gemini 3.1 Pro, last updated Sep 22 2026, still lists the model ID as `gemini-3.1-pro-preview` at launch stage 'Public preview', while third-party trackers (metacto, chapter.pub, Our Code World) list plain `gemini-3.1-pro` as GA since Feb 19 2026 — this row previously asserted the GA/no-suffix reading as settled, which the vendor page contradicts, so treat the suffix as unverified. Rate card is agreed everywhere: $2/$12 at up to 200K input, $4 in / $18 out above it, cache $0.20. SUPERSEDED AS TOP TIER: Gemini 4 Argon was announced Sep 30 2026, resolving the Sep 24 'Gemini 4 coming as soon as possible' line; Argon is Fairwind-only for now, so 3.1 Pro is still the fastest broadly available Google Pro tier. The teased Gemini 3.5 Pro was never released. | Google AI Pro $19.99/mo (was Gemini Advanced) | Gemini Free limited | |
| Gemini 2.5 Pro | $1.25 | $10.00 | 1M | Legacy Pro; >200K input $2.50 in / $15 out; cache $0.125; video understanding | Google AI Pro $19.99/mo | Gemini Free limited | |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M | Legacy cheap Flash; $0.30 text/image/video ($1.00 audio) in; cache $0.03; Flash-Lite tiers below it | Google AI Pro $19.99/mo | Gemini Free limited | |
| DeepSeek | DeepSeek V4.1 Flash | $0.15 | $0.60 | 1M | REPLACES V4 Flash (GA Sep 10 2026, model name deepseek-flash); OFF-PEAK $0.15 in / $0.60 out / cache hit $0.003 — PEAK ×2 = $0.30/$1.20/$0.006 (peak = Mon-Fri 01:00-04:00 & 06:00-10:00 UTC); 552B multimodal MoE, native vision, 384K max output, MIT open weights; KV-cache memory cut to 25% of V4-Flash for long agent sessions; legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp now billed at Flash price | DeepSeek Chat Free | 5M free tokens for new users |
| DeepSeek | DeepSeek V4 Pro | $0.66 | $1.98 | 1M | V4-Pro-0813: off-peak $0.66 in / $1.98 out / cache $0.022; peak x2 = $1.32/$3.96/$0.044 — the counter-trend rise of up to 14x on Aug 13 2026 while everyone else cut. CORRECTED AGAIN Oct 11 2026: DeepSeek's live API changelog (api-docs.deepseek.com/updates, entry dated Sep 10) now reads 'In response to user demand, we have decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged' — the Sep routing-to-V4.1-Flash plan was WALKED BACK, superseding the Oct 3 revision of this row which asserted the routing was in force. The official Models & Pricing page confirms: deepseek-v4-pro is a live column at $0.66/$1.98/$0.022 off-peak, 1M context, concurrency 500 vs Flash's 2,500, no vision on Pro. V4.1-Pro has still not shipped; watch the changelog before pinning billing assumptions. | DeepSeek Chat Free | 5M free tokens for new users |
| DeepSeek | DeepSeek R1 (Legacy) | $0.55 | $2.19 | 64K | RETIRED Jul 24 2026 along with V3.2 and the deepseek-chat / deepseek-reasoner aliases — reasoning now served by V4.1 Flash thinking mode | DeepSeek Chat Free | Free with rate limits |
| SpaceXSI (was xAI) | Grok 4.7 | $2.00 | $6.00 | 500K | NEW FLAGSHIP Sep 21 2026 — shipped as a straight upgrade to Grok 4.6 at the SAME list price and same 500,000-token context class, so for buyers it is a free quality bump unless long-horizon coding and agent persistence are the bottleneck (then it is the obvious next step); cached input $0.50; ≥200K prompt bills the whole request at $4 in / $12 out; knowledge cutoff May 2026; default reasoning effort HIGH with low/medium/high/xhigh available; text+image in, text out; Responses API and Chat Completions; function calling, web search, X search, code execution; Fast variant $4/$12; third parties undercut it (CometAPI $1.60/$4.80) — an earlier revision of this row said 4.7 was unreleased and its price unannounced, which was correct on Sep 19 and stale by Sep 21 | Grok 4.8: Musk has publicly laddered the next steps (4.8 'a noticeable improvement' — a ~2.5T model whose initial training finished mid-Sep 2026 — then 4.9 'probably Astra/Fable class', Grok 5 'maybe better than anything'), but as of Oct 19 2026 NO Grok 4.8 release, model ID, price card, context window or benchmark run has been published; 4.7 remains the flagship. NOTE THE CORPORATE LAYER: xAI renamed itself SpaceXAI in July 2026 and is rebranding as SpaceXSI (SpaceX Super Intelligence) in October; on Oct 12 2026 it launched 'The X Subscription', one five-tier plan bundling Grok, Grok Bot, Cursor and X Premium (Free / Plus $8 / Premium $30 / Super $100 / Ultra $200), which replaces SuperGrok for NEW customers — API token pricing did not change with the relaunch. | The X Subscription Premium $30/mo (ladder since Oct 12 2026: Free $0 / Plus $8 / Premium $30 / Super $100 / Ultra $200); SuperGrok Heavy $300 is legacy — existing subs keep it until a notified migration | Limited on X |
| SpaceXSI (was xAI) | Grok 4.6 | $2.00 | $6.00 | 500K | Previous flagship (Aug 12 2026, on Bedrock since Aug 19) and now superseded by Grok 4.7 at identical rates — no reason to pin 4.6 on price; cached input $0.50; ≥200K prompt bills the whole request at $4 in / $12 out; long-running agents, coding, visual work | The X Subscription Premium $30/mo (was SuperGrok $30; new five-tier ladder since Oct 12 2026) | Limited on X |
| SpaceXSI (was xAI) | Grok 4.3 | $1.25 | $2.50 | 1M | Budget tier; cached input $0.20; >200K rate $2.50/$5.00 (cached $0.40); Grok 4.1 Fast sits below at $0.20/$0.50 with a 2M window, and Grok Build 0.1 (coding) is $1.00/$2.00 at 256K. Legacy Grok 3 was $3/$15 | The X Subscription Premium $30/mo (was SuperGrok $30; new five-tier ladder since Oct 12 2026) | Limited on X |
| Xiaomi | MiMo V2.6 (Flash / Pro) | $0.14 / $0.435 | $0.28 / $0.87 | 1M | NEW ENTRY Oct 3 2026 (released Sep 21 2026) — Xiaomi's open-weight MiMo V2.6 family, and the cheapest credible multimodal tier on the board: Flash is 309B total / 15B active at $0.14 in / $0.28 out per MTok with cache hits at $0.0028, Pro is $0.435/$0.87 (cache $0.0036) and Pro UltraSpeed $4.35/$8.70; MIT weights, 1M-token context with text, image, video and audio input; Flash is the efficiency checkpoint (AutomationBench 52.3 vs Pro's 53.1) while Pro tops the Artificial Analysis open-weight index at 46 — the highest any open-weight model has reached — which makes it the direct rival to DeepSeek V4.1-Flash for cheap frontier-class work; Batch halves Flash to $0.07/$0.14; served by six providers on OpenRouter. | N/A (API + open weights) | Free to self-host (MIT) |
| Meta | Muse Spark 1.3 | $1.25 | $4.25 | 1M | SUPERSEDED Oct 12 2026 by Muse Wonder ($4/$20) and by the open-weight Muse Spark Open (Oct 15, $1.25/$4.25) — Muse Spark is no longer Meta's frontier line, so keep this row only for existing integrations. Meta's previous frontier line (Muse Spark replaced Llama as the flagship in Apr 2026); standard tier $1.25/$4.25 unchanged from 1.2; ~20% fewer tool calls and ~25% fewer tokens than 1.2; CONTRIBUTOR TIER $0.10/$0.20 (~$0.10/1M blended — cheapest model in the top-5 leaderboard) and cached input slashed from $0.15 to $0.002 — but Meta may train on your prompts and outputs (the consent switch sits outside normal security tooling); AA Intelligence Index moved 57 → 61 at xhigh effort (1.2 → 1.3); max-reasoning configuration held back at launch | N/A (API + Meta AI app) | Limited free on Meta AI |
| Meta | Muse Wonder | $4.00 | $20.00 | 1M | NEW Oct 12 2026 — Meta Superintelligence Labs' first CLOSED frontier model and the successor to Muse Spark, so Meta's break from the open-weights Llama era is now complete. Standard Wonder is the only version on the pay-as-you-go Meta Model API: $4 in / $20 out / $0.20 cached per MTok — exactly Claude Opus 5.5's card and exactly what Gemini 4 Argon reverts to on Dec 1 2026, so Google is only undercutting during its launch window. ~1M-token context (1,048,576) with roughly 943K output reported; text + image + video + audio + PDF input. 'Active Memory' carries relevant context across interactions and 'Active Focus' keeps agents working when a task gets hard — Meta has published no technical description, so treat both as product labels until you test them on long runs. AA Intelligence Index 56 — #4, tied with Claude Sonnet 5.5 and behind Opus 5.5 (58); vendor SWE-Marathon 57.5% against Opus 5.5's 46.3% and GPT-6 Astra's 42.5%, plus leads on Meta's own legal/business automation charts (vendor numbers). STRONGER TWIN: Muse Wonder Full is subscription-only in Muse Code at $50 or $200/mo and is NOT on the API — the benchmark-topping configuration is not the one you can buy per token. | Muse Code $50/mo (Power) or $200/mo (Maximum) for Muse Wonder Full; Muse app $20/$100 consumer tiers | Free base version on the Muse app; no free API tier |
| Meta | Muse Spark Open (816B) | $1.25 | $4.25 | 1M | NEW Oct 15 2026 — Meta released the WEIGHTS for an 816B-parameter (42B active) natively multimodal MoE under the custom Muse Model License, and calls it the most capable US open-weight model to date; it is the cheap route to Meta's stack now that Muse Wonder is closed. AA Intelligence Index 39 (13 points behind Muse Wonder, behind GLM-5.3's 45) but #1 among open weights on AA-Omniscience (25) with the lowest hallucination rate of any open model scoring 30+ (24%); $1.25 in / $4.25 out / $0.15 cached $ per MTok on Together, Fireworks, Baseten, Nebius and OpenRouter — about $0.84 per index task, less than half GLM-5.3's $2.01 but ~3x DeepSeek V4.1 Flash's $0.27. 1M-token context. Licence is NOT Apache/MIT: free use, modification, fine-tuning, distillation and redistribution, but you must display 'Built with Muse Spark Open' in the UI or docs of any product using it — check that before shipping. | N/A (open weights + hosted APIs) | Free to self-host (Muse Model License, attribution required) |
| Meta | Llama 4 Maverick (open weights) | $0.20 | $0.60 | 1M | Still the cheap hosted open-weight option ($0.15-$0.35 across providers (DeepInfra $0.15/$0.60; pricepertoken puts the current cheapest input at $0.188 from DigitalOcean, i.e. the floor has drifted UP ~25% in 90 days while the frontier moved down)); no longer Meta's frontier model — see Muse Spark 1.3; Intelligence Index 14.5 (45th pct) shows how far the frontier has moved | N/A (open model) | Free to self-host |
| Mistral | Mistral Large 4 (preview) | $0.68 (intro) | $2.09 (intro) | 1M | NEW Oct 6 2026 — 'le Chonk': public preview on Mistral Studio, Mistral's new open-weight frontier flagship. PARAMETER COUNT IS DISPUTED and worth flagging rather than resolving: Mistral's own model listing says 675B total / 41B active, while three independent write-ups (tech-insider, miraflow, beam.ai, all Oct 2026) converge on ~1.05T total / 49B active — treat the vendor listing as authoritative and the 1T figure as unverified press briefing. 1M-token context, multimodal input, hybrid instruct + reasoning, trained on ~3,800 Grace Blackwell GPUs in Mistral's EU datacentres. Mistral's own pricing page carries it at a launch sale: $0.68 in / $2.09 out / $0.07 cached — EXACTLY HALF the original $1.36/$4.18/$0.14 card, so budget the full rate for after the sale (no end date published). AA Intelligence Index 38 on the preview at about $1.13 per index task list ($0.57 at the discount). Weights promised by end of October 2026 (Hugging Face placeholder dated Oct 31). Until GA + weights drop, Large 3 remains the stable production row | Le Chat Pro $14.99/mo | Preview on Mistral Studio |
| Mistral | Mistral Large 3 | $0.50 | $1.50 | 262K | The stable production Mistral row since Dec 2025 — superseded as flagship by Mistral Large 4 (preview, Oct 6 2026); cached input $0.05; vision + function calling; best price-to-quality in the mid tier; Medium 3.5 above it at $1.50/$7.50 (cached $0.15), Small 4 below at $0.15/$0.60 (cached $0.015, 262K ctx, Apache 2.0 multimodal) — Mistral's own inference price page and docs confirm the Small 4 card; Ministral 3 edge family $0.10-$0.20, Codestral $0.30/$0.90, Devstral 2 $0.40/$2.00 | Le Chat Pro $14.99/mo | Le Chat Free limited |
| Alibaba | Qwen3.8-Max | $2.00 | $6.00 | 1M | Alibaba flagship (GA Aug 3 2026); flat rate across the full 1M context (no >200K surcharge); cache meters: implicit read $0.25, explicit write $2.50, explicit read $0.17; 2.4T MoE / 95B active, 1M ctx, native text+image+video, open weights (Apache 2.0); the Sep 2 2026 -0902 snapshot (coding + agentic-office post-training, terminal coding 11.3 → 29.0) auto-upgraded the qwen3.8-max endpoint on Sep 5 with BILLING UNCHANGED; third-party hosts undercut it (AIHubMix $1.69/$5.07 on the 0902 snapshot) | Sibling SKU: Qwen3.8 Max Prime (Sep 23 2026) is the same model family served for higher throughput and billed as a separate line — reasoning always on, video input added, same 1M context — at $4/$12 (pricepertoken, Alibaba being the cheap host) or ~$5.50/$16.50 through gateways, i.e. roughly 2x the standard card for wall-clock speed, not quality. | Token Plan discount | 1M free tokens |
| Alibaba | Qwen3.8-Flash | $0.15 | $0.47 | 262K (→1M) | Ultra-cheap multimodal MoE (live Aug 27 2026, QwenCloud); cache hit $0.016; NO peak surcharge; the official card is a flat $0.15/$0.47 from 0 to 1M tokens (Batch halves it to $0.113/$0.382); 262K native context extensible to 1M; open weights ship separately as Qwen3.8-Flash-Next (125B total / 6B active, Qwen Community 1.0 license) | Token Plan discount | 1M free tokens |
| Alibaba | Qwen3.7-Max | $2.50 | $7.50 | 1M | Previous flagship, still the cheapest route to Max-class quality — and after re-checking the vendor's own card this revision LEADS WITH ALIBABA'S INTERNATIONAL RATE of $2.50/$7.50 for 0| Token Plan discount | 1M free tokens | |
| Alibaba | Qwen-Flash | $0.05 | $0.40 | 256K | Budget tier that REPLACED Qwen-Turbo (Turbo no longer updated); $0.05/$0.40 up to 256K, rising to $0.25/$2.00 on longer inputs; mid-tier Qwen-Plus starts at $0.40/$1.20 (aggregator-reported rates — check Alibaba's own rate card before committing) | Token Plan discount | 1M free tokens |
| Zhipu (Z.ai) | GLM-5.3-Flash | $0.15 | $0.50 | 1M | PRICE ROSE Sep 9 2026 — the 50% launch promo ($0.075/$0.25) ended at 24:00 UTC+8, so z.ai now bills list: $0.15 in / $0.50 out / $0.03 cached (storage limited-time free); 320B MoE (18B active), native vision + video, MIT weights, 1M ctx; AA Index 57 ≈ Gemini 3.8 Flash at medium effort for ~1/6 the price; Terminal-Bench 2.1 84.3, within a point of Kimi K3's 85.0 on the coding board; cheapest top-tier blend on the board at $0.17/1M | Sibling SKU: GLM-5.3 FlashX sits between Flash and the flagship at $0.37/$1.25 per MTok — more than double Flash's rate for capability most integration work does not need; self-hosting FlashX is the cheaper route for volume. | GLM Coding Plan | 20M tokens on signup |
| Zhipu (Z.ai) | GLM-5.3 | $1.40 | $4.40 | 1M | Text flagship, same price as GLM-5.2/5.1; cached input $0.26; WEIGHTS SHIPPED Aug 28 2026 under a custom GLM-5.3 licence (MIT-style use/modify/sell, except model-as-a-service operators above $10B trailing revenue must pass a Z.ai security review); an earlier revision of this row said weights were still pending in mid-Sep — that was wrong; CyberGym 84.5% SOTA with 2,436 real vulnerabilities found; thinking cannot be disabled | GLM Coding Plan | 20M tokens on signup |
| ByteDance | Doubao Seed 2.1 Pro | $0.90 | $4.48 | 256K | Volcano Ark rate card (Jun 24 2026): ¥6 in / ¥30 out, cache hit ¥1.2 (~$0.18); Turbo is exactly half (¥3/¥15 ≈ $0.45/$2.24) and Doubao-Seed-Evolving, the coding/agent refresh, shares the same ¥6/¥30 card; text+image+video in; USD figures are conversions at 6.694 CNY (the Sep 22 2026 low per Wise; PBOC's reference rate that week was 6.7166) — not a published dollar price card, and other trackers show $0.88/$4.41 at a 6.80 rate, so the dollar column here drifts with FX rather than with ByteDance; supersedes the old Doubao 1.5 Pro (32K, ¥0.8/¥2) | Free tier on Doubao app | Free quotas on Volcano Engine |
| Moonshot | Kimi K3 | $3.00 | $15.00 | 1M | Most powerful open-weights model by GPQA (93.5%), released Jul 16 2026 with weights in late Jul under Modified MIT; 2.8T MoE with native vision; FLAT $3/$15 across the full 1M window — no long-context surcharge; cache hit $0.30; coding sibling K2.7 Code $0.95/$4.00 (cache $0.19); Batch not available on K3; CODING REFRESH Sep 11 2026 — Kimi K2.8 Preview rolled out across Kimi Code and Kimi Work under the unchanged kimi-for-coding model ID (1M ctx on all tiers, low/high/max thinking with max as default, native vision, performance described as close to K3, no published rate yet) | Kimi free app; $1 min top-up on API | No free API tier |
| 01.AI | Yi-Lightning | $0.14 | $0.14 | 128K | DISCONTINUED API — 01.AI announced on Aug 10 2026 that its open platform would progressively stop online experience, API calls and top-ups as the company refocuses on enterprise AI: new registrations/recharges stopped immediately, model experience and API service ended 2026-09-03 24:00 UTC+8, and the balance-refund window stays open to 2026-12-03 24:00. An earlier revision of this row called that shutdown unverifiable — it was wrong: the Aug 10 notice is on 01.AI's own channels and was carried by Phoenix Tech (ifeng) and Baidu Baike. 01.AI stopped pre-training large models in Mar 2025 and now sells enterprise agent platforms (万策/万智). The $0.14/$0.14 rate below is historical CAVEAT (Oct 3 2026 re-check): this pass could not independently re-confirm the Sep 3 2026 shutdown date from a first-party 01.AI page — 01.ai itself is still live and pitching enterprise/sovereign deployments, which fits the described pivot but does not evidence the date. Treat the shutdown as reported-but-not-re-verified and mark the $0.14/$0.14 rate as historical. | N/A — service discontinued | N/A (service ended Sep 3 2026) |
| Inception Labs | Mercury 2 | $0.25 | $0.75 | 128K | Diffusion (non-autoregressive) decoder, not a transformer-with-next-token — 1,009+ tok/s on Blackwell; cached input $0.025; native tool use, schema-aligned JSON, OpenAI-compatible (switch is a base-URL change); live since Mar 4 2026, #18/134 on AA intelligence; Mercury 2.5 is now the diffusion flagship at $0.20/$0.75 list (80% launch promo $0.04/$0.15), 260K context, 1,107 tok/s and roughly 40% more intelligence — pitched against GPT-5.6 Luna (low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5 | N/A | 100M free tokens on signup (raised from 10M) + 10x higher free-tier rate limits |
| TypeSafe AI | Jev (System One decision model) | $0.042 | $0.00 | 64K (32K state + question) | NEW Sep 15 2026 — a DIFFERENT SHAPE OF MODEL, not a text generator: Jev takes application state plus a TYPED question and returns a probability-weighted decision your code consumes directly, with no parsing or validation step (Kahneman System 1 vs System 2; named for economist William Stanley Jevons). Pricing is the strangest on this board — $0.042 per million INPUT tokens and OUTPUT TOKENS FREE, roughly $0.042/1M all-in, about a fifth of GPT-6 Luna's blended ~$0.20/1M (3:1 input:output) and the cheapest row here. TypeSafe claims 20-200x faster and 40-400x cheaper than frontier models on decision-shaped work, with 70-500ms end-to-end latency against 3-329 seconds for frontier LLMs. CAVEATS WORTH THE $0.042: those gains are VENDOR-RUN results, and the company itself says the headline figures are the upper end of what to expect and that it cannot prove the pricing is not subsidised (it expects prices to fall, not rise). Out of stealth with $40M seed; listed on OpenRouter around Sep 18 (Jev 1.13 dated Sep 24 2026 on model trackers); founders include ex-OpenAI researcher Diogo Almeida, and on Oct 2 2026 the CEO said Jev had reached ~25% of the Fortune 500 and was serving around a trillion tokens a day. RE-VERIFIED Oct 9 2026: price unchanged at $0.042 and jev-latest still resolves to jev-1.13.0, but the limits are now documented — 100,000 tokens/sec and 80 requests/sec, adjusting dynamically with capacity, 64K tokens per request of which 32K is state — and console.typesafe.ai issues keys without a waitlist step (one later review still describes a waitlist, so treat access as self-serve-with-limits rather than fully open). Not directly comparable to the per-token rows on this board — it is only cheaper than them for work that can be expressed as a decision, and that separations like this are why per-token price is a poor guide to cost per task | N/A (pay-as-you-go, no plan tiers) | No free tier — self-serve key from console.typesafe.ai (waitlist reportedly removed by Oct 9 2026) |
| MiniMax | MiniMax M3 | $0.30 | $1.20 | 1M | Open-weight omni-modal MoE (428B total / ~23B active, MiniMax Sparse Attention) and the best-value Chinese flagship few boards quote: the $0.30/$1.20 card is a PERMANENT 50% cut from the $0.60/$2.40 list, cache reads $0.06, and requests above 512K input tokens bill at double ($0.60/$2.40). 1M context with native image + video input; MiniMax claims parity with Claude Sonnet 4.6 on agentic benchmarks; M3 attention cuts per-token compute at 1M context to about 1/20 of the prior generation | Follow-on: MiniMax M3.1-Flash-Preview (Sep 27 2026) is preview/subscription-only (M Plan, 50%-off first month through Oct 14) with NO published pay-as-you-go rate yet — not comparable until it prices | N/A (API + MiniMax app) | Trial credits vary by host |
| Meituan | LongCat 2.0 | $0.30 | $1.20 | 1.0M | Meituan's MIT-licensed sparse MoE — re-verified Oct 3 2026 and stronger than this row previously claimed: 1.6T total parameters with 48B active, released Jul 20 2026, 1,048,756-token context and up to 262,144 completion tokens, tool calling supported (structured-output enforcement is not). OpenRouter carries it at $0.30 in / $1.20 out per MTok, which matches the limited-time promo Meituan runs alongside a 'Token Pack' flash sale and free context-cache hits; the standard pay-as-you-go card for non-cache hits is $0.75/$2.95, so $1.20 is NOT the durable output price — budget on the $0.75/$2.95 card and treat $0.30/$1.20 as promotional. MIT weights mean self-hosting is the real hedge against the pricing uncertainty; it trained entirely on Chinese chips and has led OpenRouter usage. | N/A (API + open weights) | Free tier via Meituan; free to self-host |
| Unbiased | Pareto 26.10 Preview | $0.80 | $3.20 | 1M | UPDATED Oct 3 2026 — the composite line moved on: the current release is Pareto 26.10 Preview, listed Oct 1 2026 on OpenRouter as unbiased/pareto-26.10-preview (and reachable through Unbiased's own OpenAI- and Anthropic-compatible API under the MUTABLE model string `pareto`) at $0.80 in / $3.20 out per MTok with prompt caching at $0.25 — a 3x cut from the $2.50/$7.50 that the Sep 17 listing (Pareto 26.9) carried. 1,048,576-token context, up to 131,072 output tokens, text+image in, tool calling; preliminary Oct 1 vendor numbers 69.9 DeepSWE v1.1, 50.8 Terminal-Bench 4.0, 92.4 GPQA-Diamond (denominators unconfirmed). Still a blended composite that routes across several vendors' models on every request and can change composition between releases, so no benchmark score belongs to one model — and because the API string is mutable, pin the dated preview ID if you need reproducible behaviour. | Prepaid credits | Free while anonymous (Pareto 26.9, ended Sep 16-17 2026) |
| Sakana AI | Fugu Max | $2.00 | $6.00 | — | NEW Sep 11 2026 — multi-agent ORCHESTRATION product rather than a single model (OpenAI-compatible API); cost tier of the Fugu line, undercutting Sonnet 5 / GPT-6 Sol / Kimi K3 output by 40-60% at the same $6 output as Qwen3.8-Max and Grok 4.7; capability twin Fugu Ultra v2.0 is $5/$30 (cached $0.50; >272K input $10/$45/$1.00) — the price war has moved from per-model inference to orchestration economics; web_search and web_fetch bill $0.007 per call; subscription tiers $20 / $100 / $200 per month (all include Fugu, Fugu Ultra and Fugu Max) | N/A | No free tier reported |