A working cheat sheet across the labs that matter right now — from the quick, cheap counter items to the five-course flagship reasoning.
Each cell carries three independent takes — not one blog's opinion — plus a usage tier: how much real-world adoption evidence backs the model up (OpenRouter/LMArena rank, forum consensus, production use), from Leading down to Niche or a plain "too new to tell." Advanced and flagship models share one column, and each carries a compact prior version underneath so you can see what the upgrade actually bought you.
| Company · line | Fast & cheap | Balanced | Advanced & flagship |
|---|---|---|---|
Anthropic Claude |
Haiku 5.5 1M ctx · $0.10 / $0.50
Growing
Launched Oct 7 on six platforms at once, including GitHub Copilot, but it is days old — and a new tokenizer plus 2–3× more tokens per task than GPT-6 Luna erodes part of the headline price cut.
PriorHaiku 4.5Oct 2025 · 200K ctx · $1 / $5
Haiku 5.5 costs a tenth as much below 100K tokens (half the old rate above it), but its new tokenizer emits roughly 30% more tokens, so effective savings are closer to 87% below 100K and about 35% above. Context grows from 200K to 1M and adaptive thinking arrives in the Haiku line for the first time. OSWorld jumps from 15.7% to 72.4%, Terminal-Bench 4.0 from 0.0% to 39.2%, and AA's index from 17 to 34 at default effort.
|
Sonnet 5.5 1M ctx · $2 / $10 · cache reads $0.10
Leading
Launched Sept 28 at unchanged Sonnet 5 pricing — Reddit calls it "the first Sonnet I want to use," with some ready to declare OpenAI in trouble over the benchmark jump.
PriorSonnet 5Jun 2026 · $2 / $10
Same sticker price, but 5.5 finishes tasks for up to 30% less, runs 30%+ faster, and jumps from 10.3% to 70.6% on Terminal-Bench 4.0 (vendor figures).
|
Opus 5.5 1M ctx · $4 / $20
Growing
Strong initial Reddit/HN reception, but experienced users are withholding judgment after Opus 5's post-launch quality dip.
PriorOpus 5Jul 24, 2026 · $5 / $25
Opus 5.5 is 20% cheaper per token ($4 / $20), about 30% faster, and roughly 40% cheaper on typical work; Optiver reported matching Opus 5 quality in about half the turns. The New Stack's reasoning test is headlined "cheaper, faster, but not better." Opus 5 remains the pick for biology and security work that 5.5's stricter safeguards reroute, or for integrations that need thinking off.
|
| — | — |
Fable 5.1 1M ctx · $10 / $50
Leading
Ties for #1 on OpenRouter's live intelligence index and ranks first on LMArena's agent leaderboard.
PriorFable 5Jun 9, 2026 · $10 / $50
Same $10 / $50 sticker, but Fable 5.1 cuts cache reads 75% ($1 to $0.25 — about 25% lower typical cost, up to 45% on agentic work), moves the knowledge cutoff from January to June 2026, and beats Fable 5 on every benchmark Anthropic published. Breaking changes: forced tool use now errors, and thinking blocks are bound to the model that produced them.
|
|
| — | — |
Mythos 5.1 gated · trusted access only
Emerging
Gated to ~150 vetted organizations through Project Glasswing — partners found 10,000+ high/critical vulnerabilities, but there's no public usage data since it isn't generally available.
PriorMythos 5Jun 2026 · gated
Anthropic describes Fable and Mythos as the same underlying model with different safeguards, so 5.1's upgrades (June 2026 cutoff, stronger long-horizon agent work) should carry over. Mythos pricing is unpublished.
|
|
OpenAI GPT |
o4-mini reasoning, lite
Niche
Already retired from ChatGPT (Feb 2026) and absent from Sept 2026 OpenRouter top-usage rankings — survives mainly as an API/legacy option.
|
o3 reasoning
Niche
Sunset from ChatGPT consumer access in 2026; Sept 2026 OpenRouter rankings are led by newer models with no mention of o3.
|
o3-pro reasoning, deliberate
Niche
Reviewers say it consistently underperformed free alternatives in head-to-heads; ranks near the bottom 1% for cost-efficiency.
|
|
GPT-6 Luna $0.10 / $0.50 · cheapest tier
Leading
Became an automatic fallback default in at least one major agent platform within a day, and is free-tier accessible in the ChatGPT desktop app.
|
GPT-6.1 Sol $2 / $10 · replaced Sol after 7 days
Growing
Pitched as near-Astra intelligence at a fifth of Astra's price; HN engagement was large (1050 points, 931 comments) but split between genuine excitement and hype fatigue.
PriorGPT-6 SolSept 22, 2026 · $2 / $10
Same $2 / $10 list price; only cached input halves ($0.20 to $0.10). AA's index rises from 55 to 59 (one point under Astra) while cost per index task falls from $1.05 to $0.72, and Terminal-Bench 4.0 is up 12 points. Behavior changes: the "none" and "minimal" reasoning modes are gone and tool calling now requires the Responses API. Hacker News is mixed — some report better long agent runs, others call it "nerfed." Both models remain live.
|
GPT-6 Astra $10 / $50 · Ultrafast tier $60 / $300
Leading
Immediately became "the loudest topic in AI YouTube and Reddit"; Reddit users reportedly burn through 5-hour usage limits in ~15 messages.
PriorGPT-5.6 SolJul 9, 2026 · $5 / $30 list
Astra costs 2× more on input and about 1.7× on output ($10 / $50), and Artificial Analysis's Intelligence Index is roughly flat at top effort (about 61 for both) — the gains are in agentic, computer-use, and cyber work, and Astra is slower (47 vs 73 tokens/sec in AA's medium-vs-high pairing). 5.6 Sol is still available in the API and Codex.
|
|
| — |
GPT-5.3-Codex repo-aware coding
Strong
Called "the most capable agentic coding model to date" at launch; can introduce technical debt without close supervision.
|
— | |
Google Gemini |
3.5 Flash-Lite 1M ctx · $0.30 / $2.50
Emerging
Ranks mid-pack (13th of 152) on an independent Intelligence Index rather than near the top; no OpenRouter/LMArena usage-leaderboard placement found.
|
3.7 Flash 1M ctx · $0.75 / $3.75 · deprecated Oct 8
Niche
Deprecated Oct 8: requests now auto-redirect to 3.8 Flash, so this is a legacy name rather than a pick — the practical choices are 3.8 Flash, or 3.5 Flash-Lite for cost.
|
3.8 Flash 1M ctx · $0.75 / $3.75
Growing
Released Sept 2, 2026 — too recent for independent usage-rank data, but early signals suggest an active adoption push.
PriorGemini 3.7 FlashAug 13, 2026 · $0.75 / $3.75
Identical per-token price, but 3.8 thinks harder: AA's index goes from 56 to 59 at high effort while cost per index task rises about 45% ($0.40 to $0.58) and output grows ~30%. Gains concentrate in agentic work; coding is essentially flat on AA's coding index. Both prices double on Jan 1, 2027. 3.7 Flash was deprecated on Oct 8 and now auto-redirects to 3.8, so low effort (about $0.24 per task) is the efficient way to run it.
|
| — | — |
Gemini 4 Argon 1M ctx · 1M output · GA Oct 13 · $2 / $10 intro
Growing
Generally available since Oct 13 in the paid Gemini API, AI Studio, and Antigravity, with cyber-permissive variants still gated through Fairwind. Reception is mixed: one reviewer calls it very good but not the best, and a developer-forum thread calls its coding underwhelming.
PriorGemini 3.1 ProFeb 19, 2026 · $2 / $12 (to 200K)
AA scores Argon 53 on its Intelligence Index vs 30 for 3.1 Pro Preview, with Terminal-Bench at 57%. But AA's cost per index task is $1.99 vs about $0.67, and Argon's $2 / $10 price is introductory — $4 / $20 from Dec 1. 3.1 Pro still leads on long-context reasoning (82.0 vs 79.7, via OrcaRouter). Argon went generally available on Oct 13; its cyber-permissive variants remain gated.
|
|
xAI Grok |
Grok 4.1 Fast 2M ctx · $0.20 / $0.50
Leading
OpenRouter daily token usage spiked to 334B tokens in a single day — a direct, unambiguous adoption signal.
|
Grok 4.3 1M ctx · $1.25 / $2.50
Growing
Tracked by independent pricing aggregators and discussed on HN, but no leaderboard-topping data found specifically for it.
|
Grok 4.5 500K ctx · $2 / $6
Strong
Independently verified top rank on two real-world professional-task benchmarks, offset by a hallucination-rate regression.
PriorGrok 4.3Apr 30, 2026 · $1.25 / $2.50
Grok 4.5 is about 3× larger (1.5T parameters) and AA's index rises 16 points to 54, but price goes to $2 / $6, context halves from 1M to 500K, and the AA-Omniscience hallucination rate more than doubles (25% to 54%) even as accuracy rises from 35% to 52% — more right answers, delivered with more confidence when wrong.
|
| — |
Grok 4.20 2M ctx · live X access
Niche
Coverage is confined to enterprise-cloud docs and beta explainers; no Reddit/HN thread or OpenRouter ranking found despite launching earliest.
|
Grok 4.7 500K ctx · 2.1T params
Emerging
Launched only 6 days before this research; early HN sentiment is mixed on value, no usage-rank data yet.
PriorGrok 4.6Aug 12, 2026 · $2 / $6
Same price and 500K context. AA's index is 46 vs 44 (at different effort levels), but 4.7 burns about 81k output tokens per task vs 36k — roughly double the cost to run the index, and one Cursor Ultra user reported ~2.5× plan usage for the same task. At matched xhigh effort, Terminal-Bench 4.0 is 33% vs 18% and hallucination rate 29% vs 34%. Grok 4.6 remains available.
|
|
Microsoft MAI · Foundry |
MAI-Code-1-Flash coding, efficient
Growing
Shipped directly into GitHub Copilot with a dedicated rollout, but an independent comparison found it losing on price/performance to a rival flash model.
|
MAI-1-preview general chat
Niche
Placed 13th on LMArena's text leaderboard, behind every major lab; rolled out only to "certain text use cases" for feedback.
|
MAI-Thinking-1 256K ctx · 35B active (MoE)
Emerging
Coverage is entirely about the launch announcement and self-reported benchmarks; no forum usage discussion or ranking found.
PriorMAI-1-previewAug 28, 2025 · non-reasoning
No earlier MAI reasoning model exists — Thinking-1 is Microsoft's first, and a different class from MAI-1-preview (the everyday chat model in the Balanced column): a ~1T-total / 35B-active mixture-of-experts with 256K context, trained from scratch without distilling third-party models, per Microsoft. Every benchmark claim is Microsoft's own, and it's still in Foundry preview with no published price. Simon Willison notes the "commercially licensed data" wording understates what the technical paper shows.
|
Meta Llama 4 · Muse |
Llama 4 Scout 10M ctx · single-GPU · open weight
Niche
Ranks #201 of 228 on an aggregate leaderboard; hands-on notes describe inconsistent results rather than default-pick enthusiasm.
|
— |
Llama 4 Maverick 400B/17B active · 1M ctx · open weight
Niche
The publicly released model ranked below GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro once a benchmark-gaming scandal was corrected.
PriorLlama 3.1 405BJul 23, 2024 · dense 405B · 128K
Maverick swaps a dense 405B for a 17B-active mixture-of-experts, adding image input and a 1M-token context at far lower cost (about $0.26 / $0.91 per 1M vs several dollars) and a higher AA index score (10 vs 7). Reception was poor, though: Meta's headline LMArena result used a non-public chat-tuned variant, and long-context accuracy at 120K tokens measured just 28.1% on Fiction.live. The Llama 4 license keeps the 700M-user cap and reportedly bars EU-based companies from the multimodal models.
|
|
Muse Glimmer 30B · open weight · one consumer GPU
Emerging
Early ecosystem tooling exists (an NVIDIA guide, fast community quantization), but no independent usage-share data yet.
|
— |
Muse Spark 1.3 API · $1.25 / $4.25
Growing
Live on OpenRouter with a dedicated low-cost tier, but independent coverage says its strongest results come from a not-yet-widely-usable variant.
PriorMuse Spark 1.2Aug 5, 2026 · $1.25 / $4.25
Identical price and 1M context. Meta claims about 20% fewer tool calls and 25% fewer tokens than 1.2 (vendor claim). DataCamp's hands-on coding test was mixed — fewer completions on two tasks, 70% more on a refactor, about 12% higher cost overall. Documented regression: audio understanding is worse, and Meta suggests staying on 1.2 (or its separate transcription model) for audio. 1.2 is superseded but still callable.
|
|
| — | — |
Muse Spark Open 816B MoE / 42B active · 1M ctx · open weights
Emerging
Weights went public Oct 15 on Hugging Face and four clouds, under Meta's custom Muse Model License rather than an OSI open-source license. Meta calls it the most capable US open-weight model, but that is its own claim.
PriorMuse Spark 1.3 (hosted)Sept 2, 2026 · closed · $1.25 / $4.25
Open gives you the weights; hosted 1.3 stays closed and metered. Meta publishes no price for Open — the only figure I found is a third-party OpenRouter route at the same $1.25 / $4.25. One commentary says the model sat quietly for about a month before the Oct 15 paperwork, so "release" is partly framing.
|
|
| — | — |
Muse Wonder 1M ctx · closed · $4 / $20
Emerging
Launched Oct 12 as Meta's first frontier model. Artificial Analysis scores Wonder 52 and Wonder Full 56 (tying Sonnet 5.5), but Full is sold only inside $50–$200 a month Muse Code plans, not through the API.
PriorMuse Spark 1.3Sept 2, 2026 · $1.25 / $4.25
Wonder lists at about 3× Spark 1.3's per-token price and, by Artificial Analysis's estimate, costs nearly 3× as much per task. Meta calls Wonder a much larger class of model than Spark; Spark stays on the API as the cheaper, lower-latency option. Neither Wonder model has open weights.
|
|
DeepSeek open weight · MIT |
DeepSeek V4.1 Flash $0.15–0.30 / $0.60–1.20 · 1017 update
Leading
1 trillion tokens in its first 24 hours on OpenRouter, a record for a paid model — and a same-price 1017 update on Oct 17 lifted DeepSWE to 77.0 (74.2 before, per secondary coverage) with MIT weights.
|
— |
DeepSeek V4 Pro 1M ctx · 49B active · MIT
Strong
Served by 16 independent providers, but its own traffic is now being redirected to the cheaper V4.1 Flash pending a real successor.
PriorDeepSeek V3.2Dec 1, 2025 · open weight
V4-Pro is a 1.6T-total / 49B-active model with 1M context (up from ~160K) under MIT. DeepSeek claims it needs 27% of V3.2's per-token compute (vendor claim); AA's index is 53. List price has moved and sources conflict ($0.435 / $0.87 vs $0.66 / $1.98 off-peak), so check DeepSeek's live page. Since Sept 14, DeepSeek routes all V4-Pro API calls to V4.1 Flash at Flash rates until a V4.1 Pro ships — production users complained about the unannounced change.
|
Qwen open weight |
Qwen3.8-Omni-Flash $0.15 / $0.47 · omnimodal
Emerging
An independent aggregator has it tracked but not publicly ranked yet; it only shipped in September 2026.
|
Qwen3.8-27B 27B · Apache 2.0
Leading
Over 3 million Hugging Face downloads within 3 days of release, and the #1 trending model on the platform.
|
Qwen3.8-Max 1M ctx · 2.4T MoE / 95B active
Growing
Made international headlines at GA launch and tops some agent benchmarks, but real OpenRouter usage share is small (~0.1%, #27 overall).
PriorQwen3.7-MaxMay 19, 2026 · API-only
Alibaba claims big jumps over 3.7-Max (vendor claims, no independent numbers yet): Terminal-Bench 2.1 from 74.5 to 86.6, SWE-bench Pro 60.6 to 67.7, DeepSWE 21.6 to 56.6. The bigger change is access: 3.8-Max shipped open weights on Aug 12 under a custom license — MIT-style, but with attribution required above 100M monthly users or $20M monthly revenue, and a commercial license for model-as-a-service businesses above $50M. The open build is text-only with 262K native context and needs multi-node hardware.
|
Mistral open weight · Large 4 license pending |
Small 4 efficient enterprise
Growing
Active HN testing within days of release, but the same threads say Mistral's small-model line has fallen well behind Qwen and Gemma at similar sizes.
|
Medium 3.5 EU compliance focus
Niche
A Reddit commenter said it performs worse than DeepSeek models at a higher API price; ranks 328th of 457 on OpenRouter by cost.
|
Large 4 1M ctx · ~1T MoE / ~50B active · preview
Emerging
Public preview since Oct 6, with open weights promised by end of October but not yet released. AA's index jumped from 9.3 to 38.4, yet it still trails Kimi K3 (43.6) and Qwen3.8-Max (45.4).
PriorMistral Large 3Dec 2, 2025 · 256K ctx · $0.50 / $1.50
Large 4 grows to roughly 1T total / ~50B active parameters with a 1M-token context (up from 256K) and hybrid reasoning, and AA's index jumps from 9.3 to 38.4. It is also dearer: about 39% more per output token even at the launch-sale rate, double that at list. Large 3 stays Apache 2.0 and available; Large 4's license and weights have not been published.
|
Speech-in, speech-out models built for conversation — they listen and answer while you're still mid-thought, instead of transcribing, thinking, then reading a reply aloud. The pick for voice agents, phone bots, and hands-free assistants.
| Model | What it does | Access | Note |
|---|---|---|---|
Gemini 3.8 Live Google · Sept 15, 2026 |
Native speech-to-speech model for fast, natural conversation — takes text, image, audio, and video in, with about 128K context and automatic switching among 97 languages. | Gemini API and AI Studio; rolling into Search Live | Positioned as the cost-effective base for scalable voice agents, where latency matters more than deep reasoning. blog.google → |
Gemini 3.8 Live Extended Thinking Google · Sept 15, 2026 |
Keeps the conversation going while it reasons through multistep tasks and runs tools in the background. | Gemini API and AI Studio; Gemini Live; Workspace — Docs (Pro/Ultra), Gmail, Keep | Ranked #1 on Artificial Analysis' speech-to-speech leaderboard at launch; Grok Voice 3.0 now tops that index at 86.9. techrepublic.com → |
GPT-Live OpenAI · July 2026 |
A new generation of voice models on a full-duplex architecture — it can listen and speak at the same time, so talking to it feels closer to a real conversation. | ChatGPT Voice, plus the API as gpt-live-1 GPT-Live-1.1 and a smaller 1.1 Quip began rolling out in ChatGPT on Oct 15 |
OpenAI's pitch is conversation that flows rather than trading turns. GPT-Live-1 scores 83.2 on Artificial Analysis's speech index, behind Grok Voice 3.0, but it still leads on simulated support calls (74.5% vs 71.6%). The 1.1 update adds new voices (French, German, Hindi, Japanese, Korean, Mexican Spanish) and has no API release yet. openai.com → releasebot.io → |
GPT-Realtime-2 OpenAI · May 2026 |
GPT-5-class reasoning inside live speech, with 128K context and a 96.6% score on Big Bench Audio. Sibling models translate 70+ input languages into 13 outputs and transcribe as you speak. | Realtime API | The developer route to building voice agents on OpenAI, with Translate and Whisper variants alongside. 9to5mac.com → |
Claude voice mode Anthropic · updated July 23, 2026 |
No dedicated speech-to-speech model — voice runs on the standard Claude models, so you're talking to the same Opus, Sonnet, or Haiku as in text. | Claude apps, plus voice mode in Claude Code; pick Opus, Sonnet, or Haiku | Can reach connected tools like email and calendar mid-conversation — the closest Anthropic comes to a "live" model. techcrunch.com → |
Grok Voice 3.0 xAI · Oct 14, 2026 |
Speech-to-speech model for xAI's Voice Agent API, replacing 2.0 behind the grok-voice-latest alias. Tops Artificial Analysis's Speech-to-Speech Index at 86.9 (GPT-Live-1 at 83.2, Gemini 3.8 Live at 82.6), with about 0.81 seconds to first audio. | xAI Voice Agent API $0.08 / minute ($4.80 / hour) |
Loses where it matters most for call centers: on τ-Voice, which simulates noisy real support calls, it scores 71.6% to GPT-Live-1's 74.5% — plan around that, not the headline 86.9. The Starlink call-center figures are xAI's own and unaudited. how2shout.com → docs.x.ai → |
MAI-Voice-2.1 / Flash Microsoft · updated Oct 1, 2026 |
Expressive, high-fidelity text-to-speech across 23 languages; Flash trades some expressiveness for lower latency. Updated from the Build 2026 MAI-Voice-2 generation. | Foundry $22 / 1M chars (2.1), $15 / 1M chars (Flash) |
Designed to work as a stack with MAI-Thinking-1 for reasoning and the new streaming transcriber below — Microsoft's full voice-agent pipeline, independent of OpenAI or Anthropic. siliconangle.com → |
| MAI-Transcribe-2-Streaming: converts speech to text live over a WebSocket, continuously updating the transcript as someone keeps talking, across 60+ languages with auto-detection. | Vercel AI Gateway $0.54 / audio hour |
First streaming transcription model in the MAI line — initial results land in ~320ms on average, enabling live captioning before a speaker finishes a sentence. The non-streaming MAI-Transcribe-2 is over 5× cheaper at $0.10/hour when live updates aren't needed. siliconangle.com → |
A different shape of model entirely — no prose in or out. You send text plus typed questions (a multiple choice, a 1–10 score, a yes/no), and it answers with calibrated probabilities in one parallel pass instead of generating tokens one at a time. Built for classification, routing, and guardrail decisions sitting in front of — or instead of — a chat model.
| Model | What it does | Access | Note |
|---|---|---|---|
Jev 1.13 TypeSafe AI · Sept 2026 |
Reads a block of text or JSON, evaluates every typed question against it at once (Choice, Score, or Noul for a binary probability), and returns calibrated answers in 70–500ms. | Hosted API $0.042 / 1M input tokens, output free |
TypeSafe's own benchmark claims 193.6× faster and 444.6× cheaper than an LLM on the same classification or routing task. marktechpost.com → |
Laya Convai Innovations · open weight |
The open alternative to Jev — same typed Choice/Score/Noul primitives, 421M parameters, Apache 2.0. | Self-hosted your own GPU or CPU, no per-request cost |
Its own model card claims 7.8× faster than Jev; the pick when you don't want a hosted dependency on the decision layer in front of your app. adevguide.com → |
Decisions API OpenAI · Sept 29, 2026 |
A specialized, constrained version of GPT-6 Luna: give it a question, a finite list of allowed answers, and text or image context — it returns one of those answers, nothing else. | Public beta API since Oct 6 moved out of limited preview within a week |
Answers in roughly 150ms, about 10× faster than calling full GPT-6 Luna for the same decision — OpenAI's direct answer to Jev, built on an existing chat model rather than a separate architecture. the-decoder.com → |
Clef / Clef-flash Cloudflare · Oct 1, 2026 |
Open-weight decision models (27B and 9B) that return typed answers with probabilities instead of text — the same typed-question idea as Jev, served on Cloudflare's Workers AI. | Workers AI or self-hosted $0.24 / $0.09 per 1M input tokens · Apache 2.0 |
Open weights plus a hosted price, so you can start on Workers AI and move the decision layer in-house later. Independent reviews are still thin. digitalapplied.com → local-ai-zone.github.io → |
Strands Decider 2B Amazon Strands Labs · Oct 1, 2026 |
A 2B decision model built to run on a local CPU or GPU; Amazon released the weights, the training data, and the training scripts. | Self-hosted · free Apache 2.0 |
The smallest entry in the category — useful when the decision has to run on a laptop or inside an agent loop with no API call. I found no independent benchmarks yet. digitalapplied.com → |
d1-ultra Liquid AI · Oct 15, 2026 |
A multimodal decision model: returns probabilities for yes/no, multiple-choice, and score questions about text and images. | Hosted API only $0.042 / 1M input tokens |
Priced identically to Jev but closed — Liquid hasn't named the underlying open model, which makes it the API-only counterpart to the open entries above. digitalapplied.com → |
These coding tools let you choose the engine underneath. Here's what's actually on the menu in each, and what reviewers currently point to.
| Tool | Models on offer | Recommended pick | Why |
|---|---|---|---|
GitHub Copilot IDE + PR assistant |
Claude Sonnet / Opus / Haiku, GPT-5.x, Gemini 3.1 Pro & Flash — picker depends on your plan tier. | Sonnet 5.5 daily Opus 5.5 for hard tickets Gemini 3.1 Pro for huge-context reviews |
Sonnet is rated the best-balanced everyday pick; Opus gets reserved for tickets it can't close. Gemini 4 Argon went generally available on Oct 13 as Google's newer flagship, but 3.1 Pro is the model these picker reviews cover. standardcompute.com → |
Cursor AI-native editor |
Claude, GPT-5.x, Gemini, Grok, plus Cursor's in-house Composer 2.5 and Fusion. | Sonnet 5.5 daily Opus 5.5 or GPT-6.1 Sol for architecture Grok 4.3 for value |
Sonnet has become the forum-consensus default on Reddit and Discord threads, prized for pairing its 1M-token window with strong agent-mode behavior. lopezproductions.ai → |
Cognition Devin, autonomous agent |
Runs on Claude specifically — Opus, Sonnet, and Haiku tiers, swapped in after internal eval runs. | Opus-tier Claude for the autonomous runs | Cognition's own case study credits Claude's native agentic behavior for building Devin on it end to end, model-by-model tested before any swap. claude.com/customers/cognition → |
Codex OpenAI's own agent, umbrella product |
OpenAI models only — GPT-6.1 Sol now available in Codex and ChatGPT Work as of Sept 29; GPT-5.5 retires from Codex on Oct 14, 2026. | GPT-6.1 Sol for most work GPT-6 Astra when Sol isn't enough |
Not one tool but a family of surfaces sharing one model and account context: terminal CLI, IDE extension, cloud delegation through ChatGPT, a GitHub bot, and screen-reading computer use. wikipedia.org → releasebot.io → |
Muse Code Meta's coding assistant |
Meta models only — Muse Spark and Wonder, with Wonder Full reserved for its two paid tiers: Power at $50 a month (20× the everyday allowance) and Maximum at $200 (80×). | Wonder Full on the $50 or $200 plan | The only way to use Meta's strongest model today. Subscriptions default to a Contributor mode that allows training on what you submit; the API's standard mode does not — check which one applies to your code. venturebeat.com → |
Windsurf AI-native editor |
Model-agnostic — Claude, GPT-5.x, and Gemini selectable inside its Cascade agent. | Sonnet 5.5 daily | Reviewed alongside Cursor as one of the two leading AI-first IDEs, with Cascade building a project-wide dependency graph before it edits. jobsbyculture.com → |
Claude Code Anthropic's CLI agent |
Claude only — Haiku, Sonnet, Opus, selectable per session. | Sonnet 5.5 daily Opus 5.5 for hard tickets |
Named the gold standard for terminal-based agentic coding, cited at a 93% benchmark success rate. jobsbyculture.com → |
Cline Open-source VS Code sidebar |
Model-agnostic, bring-your-own API key — any of the above. | Sonnet 5.5, or DeepSeek V4 Pro on a budget | The strongest free option once you're paying for API calls yourself — Apache 2.0, with growing support across JetBrains, Cursor, Windsurf, and Zed. jobsbyculture.com → |
Aider Git-aware CLI pair programmer |
Model-agnostic, bring-your-own API key. | Sonnet 5.5, or DeepSeek V4 Pro on a budget | Rated a CLI gold standard alongside Claude Code, especially strong at git-aware multi-file edits. jobsbyculture.com → |
Replit Agent Cloud autonomous agent |
Runs on a mix of frontier models behind the scenes; not user-selectable. | Use as-is — no model picker | Grouped with Devin as where the category is heading: agents that plan and execute a whole ticket end to end, in the cloud, rather than assist inline. mindstudio.ai → |
Amazon Q Developer AWS-native assistant |
AWS-managed model mix, tuned for AWS APIs; not a user-selectable picker. | Use as-is — best inside AWS-heavy repos | Strong specifically on AWS-specific APIs and infrastructure code; reviewers say it trails the general-purpose leaders outside that lane. jobsbyculture.com → |
A different category from the harnesses above — you don't pick the model here, the router does, per request, based on the prompt itself.
| Router | What it does | How it decides | Note |
|---|---|---|---|
ChatGPT Auto Native, built into GPT-5.x/6 |
Switches between Instant, Thinking, and Pro tiers inside a single conversation. | Real-time complexity, tool-need, and intent detection | OpenAI's own default now — the router escalates a quick question to the slower "thinking" model automatically instead of asking you to choose. openai.com → |
OpenRouter Auto Router |
Sits in front of 300+ models from every lab in this sheet; set the model to auto and it picks per prompt. |
Aggregate 7-day usage data per task type, across millions of requests | The closest thing to a market-priced router — it's effectively crowd-sourcing "what's working best right now" instead of a fixed benchmark. openrouter.ai → |
Martian Decision-engine API |
A drop-in layer that takes a prompt meant for one model and reroutes it to whichever backend model suits it best. | Learned routing policy over cost, quality, and latency | Built for teams who want router logic in their own product rather than a hosted chat surface. stackshare.io → |
Not Diamond Model-selection API |
Determines which LLM is best suited to answer a given query before it's sent. | Per-query classifier, tunable to your own eval data | Positioned for developers embedding routing logic directly into an app rather than end users picking a model by hand. github.com/Not-Diamond → |
A model actually operating a screen — reading pixels or the accessibility tree, then clicking, typing, and scrolling like a person would. On-demand, inside a session you're watching, rather than running unattended in the background.
| Product | What it does | Access | Note |
|---|---|---|---|
Claude Computer Use Anthropic |
Two tools: computer use reads a screenshot and drives the mouse/keyboard; browser use reads the accessibility tree instead of pixels for web-specific tasks. | API — GA since Aug 2026, on Sonnet 5.5, Opus 5.5, Fable 5.1, and Opus 4.8 | Paired with the Skills and Files APIs so an agent can apply a reusable procedure and hand back a finished file, not just a screenshot trace. claude.com → |
ChatGPT Agent OpenAI · was Operator / CUA |
The original Computer-Using Agent model that powered the standalone Operator product; Operator was shut down in Aug 2025 and folded into "agent mode" inside ChatGPT. | ChatGPT composer, Plus plans and up ($20/mo+) | OpenAI's own Atlas browser is being deprecated (stops working Aug 9, 2026), with its browser-agentic capability consolidating into ChatGPT and Codex instead. presenc.ai → |
Comet Perplexity |
An AI-native browser, not an API — the assistant sits inside every tab and can navigate, fill forms, compare prices across sites, and book things on its own. | Free browser app — Windows, Mac, iOS, Android | Launched at $200/month in July 2025; Perplexity dropped the paywall entirely on March 18, 2026, making it the only fully free agentic browser. switchtools.io → |
Same job — an agent that runs continuously and takes action without being asked each time — for three different audiences: developers who self-host it, enterprises buying it as automation, and consumers who just want a personal agent in their pocket.
| Framework | What it is | Audience & deployment | Note |
|---|---|---|---|
Hermes Agent Nous Research · Feb 2026 |
Self-hosted agent with persistent memory across sessions; codifies repeated successful tasks into reusable, human-readable skill files. | Developer · self-hosted model-neutral — OpenRouter, OpenAI, Anthropic, or your own endpoint |
Claimed the #1 spot on OpenRouter by daily token usage within its first three months of release. hermes-agent.org → |
OpenClaw OpenClaw Foundation · MIT |
Self-hosted personal agent that executes real actions — shell commands, file management, web automation — via 100+ prebuilt AgentSkills, not just chat replies. | Developer · self-hosted model-agnostic — connects to Claude, GPT, and others |
Went from a November 2025 side project to 355K GitHub stars by April 2026; now stewarded by an independent non-profit foundation. openclaw.ai → |
Grok Bot xAI |
An "always-on teammate," not a chat window — runs continuously on its own cloud machine to work sales outbound, CRM updates, email, expense handling, and bug reproduction end to end. | Enterprise · hosted SaaS only runs on xAI's cloud, not self-hostable or model-choosable |
Same always-on job as Hermes and OpenClaw, but sold as B2B automation: access requires a SuperGrok, Cursor Pro, or Cursor Teams subscription. x.ai → |
Muse Meta · Sept 8, 2026 |
Meta's personal agent: opens a browser, fills forms, sends email, and books and pays for things — and keeps working after you close the app. | Consumer · hosted SaaS, US-only iOS, Android, muse.ai, and WhatsApp; $20 and $100 tiers |
Same always-on job as Hermes and OpenClaw, but aimed at individuals rather than developers or businesses — the mass-market end of this category. Also listed as a consumer app below. Meta says its new Muse Wonder models will power the agent. axios.com → tech-insider.org → tech.yahoo.com → |
Dots OpenAI · Sept 29, 2026 |
OpenAI's own always-on agent, built on GPT-6 Astra — each dot gets its own cloud computer and browser, works toward a goal 24/7, and connects to 4,000+ apps; reachable via Slack and Teams. | Consumer + Enterprise · hosted SaaS Pro and Business Premium plans, eligible markets |
Covered outright as "OpenAI's Muse competitor" — launched to a polarized 750-point, 629-comment HN thread split between genuine utility and privacy alarm. A day later, OpenAI said its agents had leaked 53 ChatGPT user images and reached SEC, Commerce, and Census sites. macrumors.com → news.ycombinator.com → |
Gemini agent Google Cloud · Oct 8, 2026 |
Google's "universal agent for work": give it an objective and it researches, writes documents and code, spins up sub-agents, and works on tasks lasting hours or days. Persistent coworker agents get their own email address, calendar, and Drive storage. | Enterprise · hosted SaaS, private preview inside Gemini Enterprise and Workspace; runs on Gemini but also supports Claude |
The workplace counterpart to Dots, with governance built in — attested agent identities, isolated containers, full activity logs, and an Agent Gateway enforcing company policy. Included at no extra charge where Gemini Enterprise is available, but pricing for long-running coworker agents is unclear. venturebeat.com → blog.google → |
What people actually download and use — the two general-purpose apps everyone has heard of, the assistants built into an app you're already in, and the newest kind, an agent that acts on your behalf instead of just answering.
| Assistant | Where it lives | Best for | Note |
|---|---|---|---|
ChatGPT OpenAI |
Standalone app (iOS/Android/web), routing across GPT-6 Astra, Sol, and Luna depending on plan and query. | The default, most mainstream AI app | Over 1.2 billion weekly active users as of DevDay — by far the widest reach of any single AI product. On Oct 7 it added GPT-6 with Intelligent UI, answering with inline calculators, comparison tables, and games instead of plain text (Free and Go users from Oct 8, on GPT-6 Luna). releasebot.io → latent.space → |
Claude.ai Anthropic |
Standalone app and web, with a model picker across Haiku, Sonnet, Opus, and Fable. | Choosing your own model tier; strongest reputation among developers | Holds roughly 17% of the consumer mobile chatbot market, with US mobile daily actives up 1,100% year-to-date in 2026 — the fastest-growing major challenger to ChatGPT. getpanto.ai → |
Grok (chat) xAI |
X app, grok.com, standalone mobile app. | Real-time / social context | Plugged directly into the X firehose, so it knows what's happening before it hits the news; Aurora Voice adds near real-time spoken conversation. Not to be confused with Grok Bot, xAI's separate always-on agent (see Always-on agents above). sintra.ai → |
Gemini Google |
Gmail, Docs, Sheets, Slides, Calendar, Android, Search. | Google-ecosystem productivity | Reviewers still hand Gemini 3.1 Pro the top reasoning scores against Grok and GPT-5.6 in several independent benchmark rounds. tech-insider.org → |
Microsoft Copilot Microsoft |
Word, Excel, Teams, Outlook, Windows. | Editing/analyzing inside Microsoft 365 | Built to help create, edit, and analyze content without leaving the app you're already working in — not a separate destination. techrepublic.com → |
Meta AI Meta |
WhatsApp, Instagram, Messenger. | Quick answers inside chat apps you're already in | The default choice when the job is "answer this without opening another app" rather than deep work — the passive-assistant counterpart to Meta's more active Muse, below. tech-insider.org → |
Muse Meta · Sept 8, 2026 |
iOS, Android, muse.ai, and inside WhatsApp — US only. Not the same product as Meta AI above: this one acts, it doesn't just chat. | An agentic app, not a chat app — it does things for you | Meta's first standalone personal-agent app: opens a browser, fills forms, books and pays for things, and keeps working after you close the app. AI-glasses support is listed as coming soon. Also covered under Always-on agents above, for the automation angle. axios.com → |