This week had exactly one headline: Google's Gemini 4 Carbon topped the Artificial Analysis Intelligence Index v5.0 with 59 points on October 28. Headlines aside, most people cannot buy it today - no API, no Gemini app entry, access limited to a small group of vetted cyber defenders through the Fairwind program. The intro price of $2/$10 has no deadline mainly because there is nowhere to place an order. The launch story belongs to our news coverage, so this article answers a more practical question: among the flagship models you can actually pay for and call right now, which one should you choose? The timing is also right - v5.0 landed on October 27 and voided nearly every score in older comparisons, so the buyable flagships deserve a fresh pass.
Pick the Ruler First: AA v5.0 Only, Old Scores Are Void
Agree on the ruler first. Artificial Analysis released Intelligence Index v5.0 on October 27: knowledge work, coding, science, and general reasoning each carry 25%, and private or hidden-answer evaluations now weigh 60%. The rebalance voided all legacy numbers: early-October reports putting Argon at 53, tied with Astra, came from the old v4.3 index - under v5.0 Argon scores 50 and Astra 57 - and the earlier claim that Kimi K3.1 Pro ranked fourth overall is void too, at 46 on v5.0.
Every score here comes from AA v5.0 alone, never mixed with Arena or other leaderboards; Google's self-reported Vals Index is the one exception, labeled as a separate vendor-reported ranking. Carbon's 59 is higher, but it is not buyable, so the purchasable ceiling is 58.
The Table: Five Buyable Flagships at a Glance
Prices follow official pages as fetched on 2026-10-30; Astra and Sol pricing comes via third-party relays of the OpenAI model pages, since OpenAI's official pricing page could not be fetched directly.
| Model | AA v5.0 Score | Input/Output ($/M tokens) | Caching | Context | Availability (2026-10-30) |
|---|---|---|---|---|---|
| Claude Opus 5.5 (Anthropic) | 58 (max tier) | $4/$20 | Cache read $0.20, writes at 1.25x input | 1M (128K output) | GA: Claude API, Bedrock, Google Cloud, Azure Foundry, Claude app, Claude Code |
| GPT-6 Astra (OpenAI) | 57 (max tier) | $10/$50 | Cache read $1.00, write $12.50 | 1,050,000 (922K input cap per request) | GA: OpenAI API (Responses API for tools), all ChatGPT Pro tiers |
| GPT-6.1 Sol (OpenAI) | 55 (max tier) | $2/$10 | Cache read $0.10, write $2.50 | 1,050,000 (same as Astra) | GA, same channels as Astra |
| Gemini 4 Argon (Google) | 50 (high tier) | $2/$10 intro, $4/$20 from Dec 1 | Cache input $0.10 ($0.20 from Dec 1), 95% discount, implicit only | 1M (up to 1M output, long-decode mode) | GA since Oct 13: Gemini API (paid), Vertex, AI Studio, Gemini app, Antigravity |
| Kimi K3.1 Pro (Moonshot) | 46 (first among open-weights) | RMB 13/68 (Highspeed 26/136), roughly $2/$10 | Cache hit RMB 1.30 | 1M | GA: Kimi open-platform API, kimi.com; weights due before Nov 4 |
Claude Opus 5.5: Score King of the Buyable Field, Bill Ceiling Too
By Anthropic's official page, Opus 5.5 shipped on September 22 at $4/$20 - down 20% from Opus 5's $5/$25 - with cache reads at $0.20, a 60% cut. Context is 1M tokens with a 128K output cap. It leads the buyable field: 58 on v5.0 (max tier), 62 atop the AA Coding Agent Index with Claude Code, plus v5.0-leading FrontierCode 2.0 (46%) and HLE-Diamond (68%), with availability across all four major clouds, the Claude app, and Claude Code. If you do not need the top score, Anthropic built its own ladder: Sonnet 5.5 arrived September 29 at half the price, Haiku 5.5 on October 7.
Two caveats. First, Opus 5.5 defaults to medium effort. On AA's numbers, the low-reasoning tier scores just 35 at $0.98 per task, while the max tier hits 58 at $10.04. Score is tied to the thinking budget - the same model, ten times the cost across tiers. That is the single most important insight of this piece: ask which effort tier your task needs before picking a flagship. Second, AA's independent wrong-answer conviction rate - confident wrong answers instead of admitting ignorance - puts Opus 5.5 at 50%, tied with Astra for the worst; Anthropic's behavioral-audit claims use a different methodology and cannot offset that.
GPT-6 Astra and 6.1 Sol: Two Points Apart, Five Times the Price
Astra, the strongest and priciest GPT-6 tier released September 3, scores 57 on v5.0 (max tier), one behind Opus 5.5, and leads the scientific sub-scores: Terminal-Bench-Science (63%) and CritPt v2 (72%). Pricing, per the OpenAI model pages as relayed by third parties, is $10/$50 with cache reads at $1.00 and writes at $12.50. Context is 1,050,000 tokens, but input caps at 922K per request, and any request above 272K input tokens bills at 2x for input/cache and 1.5x for output - the million-token window works, but not at base rates. Enterprise note: by OpenAI's own framing, Astra is the first model to reach the "Critical" cybersecurity tier of its Preparedness Framework, so permission and audit design should come early.
The value pick sits beside it. GPT-6.1 Sol launched at DevDay on September 29: $2/$10, cache reads at $0.10, 55 points - two behind Astra - positioned by OpenAI as near-Astra work at roughly a fifth of the price, with GPT-6 Sol's docs already pointing users to it. Speed carries a premium - Astra's Fast mode bills 2x, Ultrafast 6x for Pro 500 and Enterprise only - while Batch and Flex cost half. A third-party estimate of a typical agent request (20K input, half cached, 2K output) puts 1,000 calls at $210 on Astra versus $41 on Sol: cache-read pricing differs by ten times, and when an agent resends its system prompt every turn, cache pricing often decides the bill more than list price.
Gemini 4 Argon: Cheap Million-Token Context, Until November 30
Argon's timeline: announced September 30, initially reserved for Fairwind cyber defenders and Google internally, globally GA on October 13. Its best price window is now: the $2/$10 intro runs through November 30, doubling to $4/$20 on December 1. Cache input is $0.10, rising to $0.20, a 95% discount covering implicit caching only. Context is 1M, and long-decode mode pushes single outputs to one million tokens - the most generous output elasticity of the five. Inputs accept text, images, video, audio, and PDF; output is text-only. Subscriptions run through Google AI Pro at $19.99/month or Ultra at $99.99/$199.99, with app availability limited to some countries.
The capability profile is "cheap but lopsided." It scores 50 on v5.0 (high tier), and its AA-Briefcase knowledge-work Elo of 1483 is the lowest of the five here. But its wrong-answer conviction rate is 4%, the lowest in this measurement - it prefers admitting ignorance, while Opus 5.5 and Astra guess half the time. Google's self-reported Vals Index (finance, coding, legal, tax) at 68.9% is a separate vendor-reported leaderboard, never shown alongside AA numbers. Per-task cost is $7.68 at intro prices; after December 1, AA's own projection puts it above $15, and the "cheap" label needs recalculating.
Kimi K3.1 Pro: Knowledge-Work Leader, With a Note for Open-Weights Readers
Released October 24, K3.1 Pro is a post-trained K3, officially "3-trillion-class" (media and AA cite 2.8T total, 104B active). It scores 46 on v5.0, first among open-weights models - but the weights are unpublished until before November 4, so that 46 cannot be downloaded or reproduced today, and the license arrives with them. Pricing follows the RMB rate card: RMB 13 input, RMB 68 output, RMB 1.30 per cache hit, Highspeed at double - roughly $2/$10 converted. Per-task cost of $3.68 is the lowest AA has measured here.
Its trump card is knowledge work: an AA-Briefcase Elo of 1678, the top score per AA's own article (this piece cites only AA's official figure). In coding, AA measured K3.1 Pro at 67.2% on Terminal-Bench 4.0, the highest within that measurement - a benchmark with multiple conflicting readings, detailed below. On reliability, AA's independent hallucination rate is 44%, improved from K3's 53% but far from Argon's 4%; the vendor-reported Hallu Bench Hard error rate fell from 23.4% to 8.3%, with the official caveat that this does not represent everyday error rates. Moonshot also states it will not release open weights advancing the absolute frontier of offensive cyber capability; the release landed as the company courted a roughly $50 billion valuation ahead of a Hong Kong IPO.
The Cost Ledger: Per-Task Cost Is More Honest Than List Price
Ranking list prices produces wrong conclusions; ranking AA's per-task costs gets closer to the truth: K3.1 Pro $3.68, Astra $6.24, Argon $7.68 (intro), Opus 5.5 max $10.04. The unbuyable Carbon sits at $7.76 - and relayed reporting (via The Decoder) notes it takes over twice as long as its predecessor on finance and tax agent tasks: identical prices do not mean identical per-task costs.
Three counterintuitive points are worth memorizing. First, Opus 5.5's $4/$20 is not the highest list price, yet its max tier posts the highest per-task cost at $10.04 - it thinks more, and thinking burns money. Work the $0.98 low tier finishes does not need max. Second, Astra has the highest list price ($10/$50) but a lower per-task cost ($6.24) than Opus 5.5 - high unit prices do not guarantee high bills; token consumption is the real variable. Third, cache pricing is the hidden ledger of agent workloads: Sol reads at $0.10, Opus 5.5 at $0.20, Astra at $1.00 - the same system-prompt-resending agent costs ten times more in cache on Astra than on Sol.
Verdict by Use Case
- Coding agents: Opus 5.5 with Claude Code tops the Coding Agent Index at 62; for cost-sensitive coding, benchmark K3.1 Pro (67.2% on AA's Terminal-Bench 4.0 reading) first.
- Knowledge work and long Chinese documents: K3.1 Pro, with the top knowledge-work score and the lowest per-task cost at $3.68, available via API today.
- Value: GPT-6.1 Sol - two points behind Astra at $2/$10 with $0.10 cache reads, arguably the sweet spot of this entire field.
- Strongest buyable overall: Opus 5.5 at 58. Argon fits 1M-context needs before November 30; re-evaluate after the December 1 doubling.
- Small teams unsure where to start: build on Sonnet 5.5 (half Opus 5.5's price) or Sol, reserving top-tier calls for the hardest 10% of tasks.
- Waiting for Carbon? No general-availability timeline exists; choose among buyable models today and revisit when it opens.
Caveats and Confidence
Three disclosures. First, v5.0 raises private or hidden-answer evaluations to 60% to prevent overfitting public benchmarks, but private items cannot be independently audited and the methodology itself is contested - treat scores as references, not constants. Second, scores move fast: Opus 5.5 led at 58 when v5.0 published on October 27; Carbon took the crown at 59 on October 28. Numbers here are current to 2026-10-30. Third, vendor self-reports and independent measurements must stay separated: Terminal-Bench 4.0 alone circulates in four readings - Google self-reporting Carbon at 65.2%, AA re-measuring 64%, Anthropic's release table listing Opus 5.5 at 66.4%, and AA measuring K3.1 Pro at 67.2% - different harnesses make them incomparable in a straight line. One scheduling note: Argon's intro pricing ends November 30, K3.1 Pro's weights land before November 4, and Carbon's expansion carries no promises - every column above may need refreshing within a month, so re-check official pages before you budget.
FAQ
Q1: What is AA v5.0, and why do the numbers differ from what I read before? A1: Version 5 of the Artificial Analysis Intelligence Index, published October 27, with private or hidden-answer evaluations at 60% and four domains at 25% each. It is not comparable with v4.3: Argon appeared at 53 in older reports but scores 50 on v5.0, Astra 57. When numbers clash, check the index version first.
Q2: Which buyable flagship is strongest overall right now? A2: Claude Opus 5.5 at 58 (Carbon's 59 is higher but has no API). Mind the tiers: Opus 5.5 defaults to medium effort, its low tier scores 35 at $0.98 per task, and the 58 belongs to max at $10.04.
Q3: What should I pick on a tight budget? A3: For general English tasks, GPT-6.1 Sol at $2/$10 with $0.10 cache reads and 55 points. For Chinese and knowledge work, K3.1 Pro at $3.68 per task with the top knowledge-work score, live via API today.
Q4: Which model for coding agents? A4: Claude Code with Opus 5.5 tops the AA Coding Agent Index at 62; on a budget, K3.1 Pro measured 67.2% on AA's Terminal-Bench 4.0 reading at roughly half the price. Run both on your own real tasks first.
Q5: Should I wait for Carbon instead of buying now? A5: Google has published no timeline for general access; today Carbon serves only a small vetted group of cyber defenders. Ship on a buyable model now and re-evaluate when it opens - its $2/$10 intro price has no deadline, but only if there is somewhere to place the order.
Related reading
- Kimi Cut Its Cyber Attack Skills Before Open-Sourcing: Why? - K3.1 Pro's safety trimming and launch, the backdrop to its row above.
- OpenAI Ships GPT-6 Astra, Declares AGI Era Begun - the launch background behind Astra's row in the table.
- Anthropic S-1 Filing: $16.3B Revenue, $80.7B Net Loss Decoded - the financial backdrop to Opus 5.5's price cuts.