The frontier AI models, right now

Long-context summarization

Gemini 4 Argon (Google)

A unified 1,048,576-token limit shared by input, thinking and output, with text, image, video, audio and PDF input — per Google's model page for gemini-4-argon, which lists the context window as a single input-and-output figure and a maximum output of 65,536 tokens per synchronous response, or up to 1,048,576 with long decode. It is the first Gemini window where output is not capped at 64K, which matters for summarization that has to write long as well as read long. The Pro slot is no longer preview-only: Argon is Stable GA (October 13, 2026) and takes it from Gemini 3.1 Pro Preview. DeepSeek-V4.1-Flash matches the 1M input window in MIT-licensed open weights for the self-hosting path; Anthropic's Claude Fable 5.1 carries a 1 million-token context if a closed-weights alternative is preferred.

Code generation

GPT-6 Astra (OpenAI)

OpenAI calls Astra “the best model for software engineering to date” and its launch Coding table leads the comparison on Terminal-Bench 4.0 (57.9% against Claude Fable 5.1's 55.8%) and DeepSWE v1.1 (74.1% against 67.4%) — per the GPT-6 Astra announcement. Read the caveats, because this is the closest call on this page. Astra publishes no SWE-Bench figure of any variant, where its predecessor claimed SWE-Bench Pro 64.6, so the benchmark most often quoted for this task now has no OpenAI entry at all; and on the third-party Artificial Analysis Coding Agent Index in that same table Astra's 67.0 sits behind Claude Opus 5's 68.1. Among flagships, Anthropic's Mythos-class Claude Fable 5.1 (September 1, 2026) is the strongest closed-weights alternative, at 81.2% on SWE-bench Pro. Anthropic's newer, cheaper tiers now post higher Terminal-Bench 4.0 and SWE-bench Pro figures than either flagship: Claude Opus 5.5 (September 22, 2026, $4 / $20) reports 66.4% and 89.9%, and Claude Sonnet 5.5 (September 28, 2026, $2 / $10) reports 70.6% and 81.3%, per Anthropic's launch posts and system cards. Neither is Anthropic's flagship, which is the only reason neither takes this pick. Alibaba's Qwen3.8-Max (August 3, 2026) debuted fourth on Arena's Frontend Code leaderboard at launch.

Math and reasoning

GPT-6 Astra (OpenAI)

OpenAI's August 2025 GPT-5 unified router merged the o-series reasoning track into the main line, and the GPT-6 generation carries it forward on the same low–max reasoning-effort ladder. Astra's launch tables claim GPQA Diamond 96.0%, ARC-AGI-2 95.0% and FrontierMath Tier 4 (v2) 97.6%, the highest any flagship on this roster publishes on all three, and the ARC Prize Foundation independently verifies the ARC-AGI-2 score at 95.0% on its public leaderboard — per the GPT-6 Astra announcement. No current flagship publishes an AIME score any more; the labs have moved to FrontierMath and research-level substitutes.

Vision and multimodal

Gemini 4 Argon (Google)

Largest deployed multimodal surface of any frontier line: Search AI Overviews, Workspace, Pixel + Android on-device Nano, Chrome Built-in AI. Gemini 4 Argon takes text, image, video, audio, and PDF input natively, with agentic video understanding generally available and computer use in preview, per the Gemini 4 Argon model page; it is a Stable GA Pro-slot model rather than the preview that held that slot until October 13, 2026. Modality coverage for the rest of the line is on the Gemini model docs. Alibaba's Qwen3.8-Max (August 3, 2026) is the new contender — native text, image, and video input at $2.00/$6.00 per million tokens, though the open-weights build Alibaba published on August 12 is text-only and drops the vision path. GPT-6 Astra takes text and image input natively and is the OpenAI alternative if you're already on that stack.

Cost per million output tokens

DeepSeek-V4.1-Flash

MIT-licensed open weights and the lowest hosted-API output price of any flagship on this roster — $0.60 per million output tokens off-peak and $1.20 at peak — per the DeepSeek pricing page. The September 10, 2026 release cut every leg against the V4-Flash tier it replaced, and the peak/off-peak split no longer changes the answer: at $1.20 the peak rate still sits below Muse Spark 1.3's $4.25, the next-cheapest flagship on the roster now that Google's flagship is Gemini 4 Argon at $10 rather than 3.8 Flash at $3.75, so DeepSeek is the cheapest around the clock and not only off-peak. The older deepseek-v4-pro is still served, unchanged, at $1.98 output off-peak and $3.96 at peak — DeepSeek's change log says it kept it alive after September 14, 2026 in response to user demand. Flat-rate billing ended August 16, 2026, so the number still depends on when the batch runs: peak is 01:00–04:00 and 06:00–10:00 UTC on weekdays only, Chinese public holidays excepted, and everything else bills off-peak. Self-hosting drops the cost further when volume justifies the GPU spend. Several non-flagship SKUs still go lower, though the September cut closed most of the gap: Meta's muse-spark-1.3-contributor at $0.10 / $0.20 — priced on the condition that Meta may train on your prompts — and OpenAI's gpt-6-luna at $0.10 / $0.50 undercut V4.1-Flash's off-peak rate on both legs — as does Anthropic's claude-haiku-5-5 (October 7, 2026) at the same $0.10 / $0.50 for prompts up to 100K tokens, per Anthropic's pricing page — and Alibaba's qwen3.8-flash at a flat $0.15 / $0.47 undercuts it on output. Google's gemini-2.5-flash-lite also undercuts it on both legs at $0.10 / $0.40, but it is no longer a tier a new project can reach: on September 18, 2026 Google restricted the 2.5 models to accounts that had already used them and pointed new work at 3.5 Flash-Lite ($0.30 / $2.50) or 3.8 Flash instead. OpenAI's older gpt-5-nano ($0.05 / $0.40) and gpt-4.1-nano ($0.10 / $0.40) also still list under V4.1-Flash on both legs, but both are on OpenAI's shutdown schedule — gpt-4.1-nano on October 23, 2026 and the gpt-5-nano snapshot on December 11, 2026, each with gpt-5.6-luna named as its replacement. Mistral's Ministral 3 line undercuts it on output as well — 3B at $0.10 / $0.10, 8B at $0.15 / $0.15 and 14B at $0.20 / $0.20, per Mistral's pricing page — while Mistral Small 4 at $0.15 / $0.60 sits level with V4.1-Flash rather than a tier under it, matching its off-peak rate exactly. Of the current closed-weights small tiers named here that carry no training-on-your-data condition, GPT-6 Luna and Claude Haiku 5.5 are lower on input ($0.10 against $0.15) and qwen3.8-flash lower on output ($0.47 against $0.50), with Gemini 3.1 Flash-Lite at $0.25 / $1.50 and 3.5 Flash-Lite at $0.30 / $2.50 above both.

Agentic tasks (tool use, computer use)

Claude Fable 5.1

Most mature agentic ecosystem: tool use, the Model Context Protocol (MCP) and computer use, documented in the Anthropic tool-use docs, plus Claude Code's dynamic workflows, which script up to 1,000 subagents per run with 16 running at once by default, per the Claude Code docs. Claude Fable 5.1 (September 1, 2026) is the current head of that stack, announced as setting “a new standard for coding, knowledge work, and long-running problem-solving tasks” and sitting above the Opus-tier head Claude Opus 5.5 (September 22, 2026), which Anthropic's docs route the harder agentic work up from. xAI's multi-agent Heavy mode (Grok 4.20, 16-agent Heavy mode) is a different agentic shape.