GPT-6 Astra
OpenAI • Free: Paid
Context: ~1.05M in / 128K out*
Multimodal: Text + image in
Frontier, high-cost. AA Coding Index 76.9 (reported).
AI Models 2026: 26 Models Compared
GPT-6, Claude 5.5, Gemini, Grok, DeepSeek, Qwen, Kimi, GLM, MiniMax and more. Which model leads in coding, reasoning, vision and cost, and what the benchmarks do not prove. Values marked * are conflicting or approximate in sources. Re-check the primary leaderboard before you decide.
No single model leads everywhere. That is the main finding.
GPT-6 Sol
OpenAI • Free: Paid (Chat)
Context: ~1.05M / 128K out*
Multimodal: Text + image in
Balanced frontier. AA Index ~47.5-48 (max effort, reported).
GPT-6 Luna
OpenAI • Free: Free / Go (Chat)
Context: ~1.05M / 128K out*
Multimodal: Text + image in
Cheapest frontier-adjacent. $0.10 / $0.50 per 1M in/out (reported).
gpt-oss 120B / 20B
OpenAI • Free: Download
Context: Not stated in sources
Multimodal: Text only
Open reasoning models. Tool use, adjustable reasoning.
Claude Opus 5.5
Anthropic • Free: Paid
Context: 1M / 128K out
Multimodal: Text + image in
AA Index ~57.6-58 (reported leader). Strongest independent coding evidence. $4 / $20.
Claude Sonnet 5.5
Anthropic • Free: Paid
Context: 1M / 128K out
Multimodal: Text + image in
AA Index ~56 (reported). Best balance of quality + cost for many tasks. $2 / $10.
Claude Haiku 5.5
Anthropic • Free: Paid / high-volume
Context: 1M / 128K out*
Multimodal: Text + image in
Fast, high-volume. Free tier uses older Sonnet/Haiku.
Claude Fable 5.1
Anthropic • Free: Paid plans
Context: 1M / 128K out
Multimodal: Text + image in
Safeguard-restricted version of Mythos-class. Mythos restricted access.
Gemini 3.1 Pro
Google • Free: Varying Free / paid
Context: 1M*
Multimodal: Text, image, audio, video
Multimodal breadth. $2 / $12 (reported).
Gemini 3.8 Flash
Google • Free: Free app / API
Context: 1M in / 64K-65K out*
Multimodal: Multimodal
Frontier-level at Flash price. $0.75 / $3.75 promo (reported, doubles 1 Jan 2027).
Gemini 4 Argon
Google • Free: No public free tier
Context: 1M output claimed*
Multimodal: Text, image, video, audio claimed
Announced 30 Sep 2026. Fairwind defenders first. No public date. AA Index ~53 (independent).
Gemma 3 (1B-27B)
Google • Free: Free
Context: 128K (32K for 1B)
Multimodal: Text + image (4B+)
Small open multimodal. Good for research / edge.
Muse Spark 1.3
Meta • Free: Meta AI chat
Context: 1M in / up to ~943K out*
Multimodal: Text, image, audio, video
Meta flagship. Open weights promised 'soon', not shipped in sources.
Muse Glimmer 30B
Meta • Free: Download
Context: Not stated
Multimodal: Text + visual
Open 30B model from Meta.
Llama 4 Scout
Meta • Free: Download
Context: 10M
Multimodal: Text + image
Largest advertised context. Not frontier-quality leader in 2026 sources.
Llama 4 Maverick
Meta • Free: Download
Context: 1M
Multimodal: Text + image
Open multimodal. No new Llama generation shipped in sources.
Grok 4.7
xAI / SpaceXAI • Free: Via Grok / API
Context: 500K, no text-output limit*
Multimodal: Text + image
Latest Grok in sources. $2 / $6 (<200K, reported). Grok 5 not confirmed.
Grok 4.6
xAI / SpaceXAI • Free: Via Grok / API
Context: 500K
Multimodal: Text + image
Predecessor to 4.7. Available in Cursor / Grok Build / API.
DeepSeek V4.1 Flash
DeepSeek • Free: Free chat / API
Context: 1M*
Multimodal: Native image input reported
Cheap + open. ~$0.15 / $0.60 off-peak (reported).
DeepSeek V4 Pro
DeepSeek • Free: Free chat
Context: 1M
Multimodal: Text; vision-exp variant reported
Open frontier-scale. Vendor SWE runs high, harness-dependent.
Mistral Large 3
Mistral • Free: Le Chat free / API
Context: 256K
Multimodal: Text + image
Open flagship from Mistral.
Mistral Large 4 (preview)
Mistral • Free: Preview API
Context: 1M docs / 524K preview limit*
Multimodal: Text + image
1.05T total / 49B active in most coverage (52B in one source - conflict).
Qwen 3.8-Max
Alibaba • Free: API
Context: 1M / 131K out
Multimodal: Text, image, video in
2.4T / 95B (reported). $2 / $6 (reported).
Kimi K3
Moonshot • Free: Download
Context: 1M
Multimodal: Text (+ multimodal in family)
2.8T / 104B active (reported). AA Index 44 (reported). Agentic / coding focus.
GLM-5.3-Flash
Z.ai / Zhipu • Free: Download / API
Context: 1M* (conflicting 300K-1.31M)
Multimodal: Native multimodal
320B / 18B active (reported). $0.15 / $0.50 (reported). Base GLM-5.3 weights staged.
MiniMax M3
MiniMax • Free: Download
Context: 1M
Multimodal: Text + image + video
428B / 23B active (reported). $0.30 / $1.20 (reported). Licence: Apache in one source, custom in another.
Benchmarks, Who Leads Where?
Read this first: Most headline numbers are vendor-reported or secondary-reported, not independently re-run here. Scaffold / harness alone can move SWE-bench Verified by 5-15 points. Compare same effort + same harness, or do not compare. 1-2 point gaps are noise.
| Job | Leader (reported) | Evidence |
|---|---|---|
| Overall / general | Claude Opus 5.5 | Artificial Analysis Intelligence Index v4.3.2 (secondary): Opus 5.5 ~57.6-58, Sonnet 5.5 ~56, Gemini 4 Argon & GPT-6 Astra ~53, Sol ~47.5-48. |
| Coding | Claude Opus 5.5 / Claude Code (independent); GPT-6 Astra close | GPT-6 Astra AA Coding Index 76.9. SWE-bench Verified official like-for-like (Feb 2026): Opus 4.5 76.8%. Vendor / Vals runs (different harness) run 93-97%, not comparable. |
| Reasoning / math | No single winner | GPQA Diamond ~92-94% clustered (198 Qs, small-n). HLE 46-61% depending on tools / effort. Do not rank on 1-2 point GPQA gaps. |
| Long context | 1M is standard | Most frontier: 1M in. Llama 4 Scout 10M is an outlier. Gemini 4 Argon claims 1M output (not publicly verifiable). |
| Multimodal / vision | Google Gemini (breadth); Claude / GPT-6 close on visual reasoning | Gemini 3 Pro MMMU-Pro 81%, Video-MMMU 87.6% (Google). AA MMMU-Pro: Opus 5.5 87.7%, Sol 83.3%. |
| Speed / cost | Small / open models by far | Cheapest reported: Luna $0.10 / $0.50, DeepSeek V4.1 Flash ~$0.15 / $0.60 off-peak, GLM Flash $0.15 / $0.50. Top: Opus 5.5 $4 / $20, Astra $10 / $50, Fable 5.1 $10 / $50 per 1M in/out. Cost per task flips token-price rankings because output volume differs. |
Sources to re-check on publication day: LMArena, SWE-bench, Scale HLE, Artificial Analysis, Vals AI. Prices change weekly.
Free / Updated Models With Free Tiers
Free tiers exist for all major chatbots, but caps, models and ads change without notice. Verify on the provider pricing / help page.
- ChatGPT: GPT-6 Luna in Free / Go (Chat from 8 Oct 2026, reported).
- Claude: Sonnet-class + Haiku web chat, capped (~15-100 msgs / 5-hr range in different summaries).
- Gemini: Flash-class in app, compute-based limits.
- Meta AI: Muse Spark via meta.ai / WhatsApp / Instagram / Facebook, daily limit reported.
- Grok: Limited Grok on grok.com + X, unpublished limits.
- DeepSeek: V4 Flash chat, fair-use throttling at peak.
- Open weights (self-host): Gemma 3, Llama 4, Qwen 3 / 3.8, DeepSeek V4, GLM-5.3-Flash, Kimi K3, MiniMax M3, gpt-oss, Muse Glimmer. No per-use fee, but you pay compute / hosting. “Open” does not mean “free to run”, large MoEs need datacenter-class hardware.
Agentic AI in 2026
What it means: Systems that pursue a goal over multiple steps, plan, choose actions, use tools (files, APIs, browser, code execution, search), observe results, and adapt, rather than producing a single response. MCP is a widely used standard for connecting agents to tools / data.
Major tools / frameworks: Claude Code, OpenAI Codex, Cursor, Grok Build, Gemini CLI / Antigravity, Devin, Windsurf, Muse Code (coding). Manus, ChatGPT Agent mode, Claude Cowork, Grok Bot / Team Bots, Gemini Spark / Enterprise Agent Platform (general / workplace). LangGraph, CrewAI, Microsoft Agent Framework, OpenAI Agents SDK, Google ADK, Claude Agent SDK, LlamaIndex, Pydantic AI, smolagents (builders).
Can do: Write / edit / refactor code, run tests, fix failures, open PRs. Multi-step research to draft report / spreadsheet / slides. File / folder work, browser / API workflows, long-running jobs, parallel subagents.
Cannot reliably do: Long-horizon / ambiguous projects (best on 1-2 well-scoped tickets). Architecture / security by default (local fixes, hallucinated APIs, skipped validation). Irreversible / high-stakes actions without approval and audit. Novel invention. Cost / reliability vary by harness, effort and model. Open-weight agents can have safeguards removed, do not present open weights as “safe by default”.
AGI Status, Hype vs Reality
Bottom line: No lab has produced a generally accepted, independently verified demonstration that AGI has been achieved. Claims vary by definition, audience and date, and often contradict each other from the same company.
- No agreed definition. “AGI” is used to mean human-level on most professional tasks, smarter than the smartest human, Nobel-level across disciplines, or a marketing era label. Without a definition, a claim of arrival is not testable.
- What leaders have said (reported, not endorsed): Sam Altman, OpenAI “not quite yet” there (Aug 2026) while Greg Brockman wrote “Welcome to the AGI era” at GPT-6 Astra launch. Dario Amodei (Anthropic) prefers “powerful AI” over AGI. Demis Hassabis (Google DeepMind), most measured, ~50% by ~2030 in 2026 reporting. Elon Musk, shortest timelines. Yann LeCun, Fei-Fei Li, Andrew Ng, sceptical / term should be retired.
- What models actually show: Real gains in coding, long-horizon agents, math / science benchmarks and multimodal work, but still narrow, harness-dependent and error-prone. HLE top scores (~46-61% depending on setting) remain well below human expert ~90% cited in benchmark literature.
- How to write it: Say “labs disagree, definitions differ, no consensus AGI exists”, not “AGI is here” or “AGI is 10 years away” as fact. Attribute every timeline to its speaker + date.
How We Will Keep This Updated
This page is a snapshot, not a live leaderboard. Model line-ups change weekly in 2026. We will update it when a major model ships, a primary leaderboard moves, or a free tier changes, and we will keep the “* = conflicting / approximate” flags instead of hiding uncertainty.