Eight categories. Twenty-four leading models. Updated monthly. AI-friendly citations included.
One issue a month, all archived. Trace how the AI capability map shifts month over month.
June 9 — Anthropic ships Claude Fable 5, the first publicly available Mythos-class model, and it debuts at #1 on the Artificial Analysis Intelligence Index v4.0 with a score of 65, clearing Opus 4.8 by 4 points. Anthropic now holds the entire reasoning frontier.
Previously: Claude Opus 4.8
Released June 9. First public Mythos-class model — debuts #1 on the AA Intelligence Index v4.0 at 65 and scores 53% on Humanity's Last Exam (7+ points clear).
Released May 28. Adaptive reasoning + max-effort mode; now #2 on the AA Intelligence Index v4.0 behind Fable 5.
Apr 24 release. 1M-token context, native MCP + Skills + computer use + hosted shell.
Strong agentic frontier with Antigravity 2.0 platform integration.
Released Apr 30. AA Intelligence Index 53 with always-on reasoning, 1M context, and aggressive pricing — the cheapest of the frontier four.
Released June 9, 2026. AA Intelligence Index v4.0: 65 (Adaptive Reasoning, max effort) — 4 points ahead of Opus 4.8 (61.4) and GPT-5.5 xhigh (60.2); Gemini 3.1 Pro (57) and Grok 4.3 (53) follow.
Fable 5 is the Mythos-class tier above Opus, gated with extra safeguards; Opus 4.8 stays the default workhorse for cost-sensitive high-volume work.
OpenAI ships GPT Image 2 with token-based pricing and 50% Batch API discount. Recraft V4.1 holds the Artificial Analysis quality leaderboard, while Adobe Firefly enterprise mode remains the rights-cleared default.
Previously: GPT Image-2
Released April 21. State-of-the-art quality with token pricing and Batch API support.
Leads Artificial Analysis text-to-image arena on raw output quality.
IP-cleared training data; the enterprise-safe choice for commercial use.
Released April 21, 2026. Token-based pricing, Batch API at 50% off.
Recraft V4.1 leads AA's text-to-image arena on raw quality; GPT Image 2 wins on ecosystem and pricing transparency.
Seedance 2.0 (ByteDance) tops the Artificial Analysis text-to-video Arena with audio at ELO 1215 — the first model to make synced audio-visual generation state of the art. Google's Veo 3.5 leads the silent-cinematic tier; Kling 4 holds the value end. OpenAI's Sora exited the field after its app was discontinued on April 26, 2026.
Previously: Veo 3.5
Tops the Artificial Analysis text-to-video Arena (with audio) at ELO 1215 — ahead of Veo. Native audio-visual generation: 15-second multi-shot clips with synced sound from text, image, audio and video inputs.
Production-grade film output with strongest temporal coherence.
Dominant in APAC short-form ad creative; fastest iteration cycle in the space.
Seedance 2.0 holds #1 on the with-audio text-to-video Arena; Sora 2 drops out after OpenAI discontinued the Sora app (Apr 26, 2026).
Seedance 2.0 for synced audio-visual and multi-shot narrative; Veo 3.5 for cinematic fidelity; Kling 4 for cost.
Claude Fable 5 launches June 9 and immediately tops SWE-bench Pro at 80.3% — about 11 points clear of Opus 4.8 (69.2%) on the same uncontaminated multi-language benchmark. Pro scores run far below the contaminated Verified set, so treat 80% as the honest new frontier.
Previously: Claude Opus 4.8
#1 on SWE-bench Pro at 80.3% (June 9 launch) — 11 points ahead of the field. Mythos-class agentic coding for the hardest multi-file, long-horizon work.
#2 on SWE-bench Pro at 69.2% (updated June 8), behind Fable 5. The cost-sensitive default for multi-file refactoring, code review, and long-horizon agentic edits.
#3 on SWE-bench Pro (June 8). Specialized terminal-code agent; AA Intelligence Index 54.
Ranks #3 on AA Coding Agent Index. IDE-native pair programming with multi-file context.
Fable 5 debuts June 9 at #1 on SWE-bench Pro (80.3%); Opus 4.8 (69.2%) #2, GPT-5.3 Codex #3.
Fable 5 for the hardest multi-file refactors and long-horizon agents; Opus 4.8 as the cost-sensitive default; GPT-5.3 Codex for sandboxed terminal agents; Cursor Composer 2.5 for IDE-native pair programming.
OpenAI's Realtime 2 (May 7) brings configurable-reasoning speech-to-speech to general availability; AA's text-to-speech crown goes to Fun-Realtime-TTS, and MAI-Transcribe-1.5 wins STT on accuracy-speed.
Previously: ElevenLabs v3
GA on May 7. Configurable-reasoning speech-to-speech with realtime translate + Whisper variants.
Industry default for character voice cloning and audiobook production.
Tops AA text-to-speech leaderboard on quality metrics.
Realtime 2 family shipped May 7 (gpt-realtime-2 / -translate / -whisper).
Realtime 2 for agentic voice; ElevenLabs v3 for character voice cloning; Fun-Realtime-TTS for raw TTS quality; MAI-Transcribe-1.5 for transcription.
Suno v6 widens the gap on full-song coherence and lyric prosody; Udio v3 keeps pushing studio-grade mixing; Lyria (Google) integrates into Gemini Omni for any-to-music workflows.
Previously: Suno v5.5
Rolling release expected late June. Best full-song coherence and lyric prosody.
Studio-grade mixing with stem-level output for producers.
Folded into Gemini Omni for any-to-music + cross-modal generation.
Suno v6 expected late June; v5.5 remains the deployed default.
Suno v6 for full-song generation; Udio v3 for mixed studio-grade stems; Lyria via Gemini Omni for cross-modal generation.
Anthropic sweeps LMArena Vision top-3 with Opus 4.7-thinking, 4.6-thinking, and 4.7. Opus 4.8 is too freshly released to appear on Arena ELO but is expected to consolidate the lead by end of June.
Previously: GPT-4o Vision
Tops LMArena Vision. Best OCR + chart + document understanding.
Strongest image-grounded reasoning chains; native computer-use vision pipeline.
Best video understanding and long-form temporal reasoning.
LMArena Vision top-3 all Anthropic (1309 / 1303 / 1298 ELO).
Anthropic for OCR + document Q&A + chart understanding; GPT-5.5 for image-grounded reasoning; Gemini 3.1 Pro for video understanding.
Kimi K2.6 (Moonshot) leads open weights at AA Intelligence Index 54 — within 7 points of frontier closed models. DeepSeek V4 Pro (MIT, 52) is the #2 open reasoning model, and Google's new Gemma 4 12B (Apache 2.0, 2026-06-03) packs native multimodal into a 16GB-laptop footprint. The closed-vs-open gap is the narrowest it has ever been.
Previously: Llama 4
Open-weights leader on AA Intelligence Index. Closes the closed-source gap to 7 points.
MIT-licensed open weights at AA Intelligence Index 52 — #3 of 89 overall and the #2 open reasoning model behind only Kimi K2.6.
Released 2026-06-03 under Apache 2.0. Encoder-free native multimodal (text/image/audio/video), 256K context, runs on a 16GB laptop — performance nearing last-gen 27B.
Best open-source for Chinese-language self-host deployment.
Kimi K2.6 (54) tops open weights; DeepSeek V4 Pro (52) #2; Gemma 4 12B brings 16GB-laptop multimodal.
Kimi K2.6 for general-purpose open deployment; DeepSeek V4 Pro for cheap frontier-grade reasoning; Gemma 4 12B for on-device multimodal; Qwen3.7 Plus for Chinese self-host.
The 2026 value war is led by China's open-source camp. DeepSeek V4 Flash delivers near-flagship intelligence (Index 47) at roughly a tenth of the price — a blended cost near $0.06 per 1M tokens. The leaderboard makes one warning explicit: 'Flash' and 'mini' branding does not mean cheap. Gemini 3.5 Flash scores a strong 55 but costs over 20× more per full Intelligence-Index run.
The intelligence-per-dollar king. AA Intelligence Index 47 at $0.14/$0.28 per 1M tokens — about a tenth of comparable Flash flagships, with the lowest cache-hit price of any 2026 frontier model.
Best balance in the high-intelligence + low-price quadrant. Intelligence Index 52 (top-3 overall) at $0.435/$0.87 — a fraction of same-tier flagships like GPT-5.5 and Claude Opus.
Keeps the value race from being a single-vendor story. Intelligence Index 53 at $0.40 input — a higher score than V4 Pro; the trade-off is pricier output and slower generation.
New category. DeepSeek V4 Flash leads intelligence-per-dollar; open-source models sweep the value tier.
DeepSeek V4 Flash for highest-volume low-cost workloads; V4 Pro when you need stronger reasoning but still want to save; Qwen3.7 Plus to diversify away from a single vendor.
What changed across the AI model landscape this month — distilled from the data above.
Claude Fable 5 (Mythos-class, June 9) debuts #1 on the AA Intelligence Index (65) and #1 on SWE-bench Pro coding (80.3%, ~11 points clear); Opus 4.8 sits #2 on both, and Opus 4.7-thinking still holds the LMArena Vision and Document arenas. No single lab has held reasoning + vision + code this completely since GPT-4-era OpenAI.
OpenAI's April 24 GPT-5.5 ships with 1M-token context, native MCP, Skills, hosted shell, computer use, tool search, and web search — turning the API itself into an agent runtime.
May launch of Gemini Omni unifies image / audio / video generation; Antigravity 2.0 platform turns Gemini 3.5 into an agentic substrate. Google's bet: not the smartest single model, but the most integrated stack.
Kimi K2.6 (54) is within 7 points of Claude Opus 4.8 (61) on AA Intelligence Index. Meta's muse-spark cracks LMArena top-5 at 1489. Closed-source's moat is the smallest it has ever been.
Qwen3.7 Max (Alibaba, 57), MiniMax-M3 (55), Kimi K2.6 (Moonshot, 54), MiMo-V2.5-Pro (Xiaomi, 54), and Qwen3.7 Plus (Alibaba, 53) all sit in the AA Intelligence Index top 15 — Chinese labs are no longer 'catching up', they're inside the frontier.
OpenAI Realtime 2 (May 7) + Gemini 3.1 Flash TTS (April) + Fun-Realtime-TTS push real-time voice agents from research to production. Speech-to-speech with reasoning is now a checkbox feature.
DeepSeek V4 Flash delivers Intelligence Index 47 at a blended ~$0.06 per 1M tokens — open-weights models from DeepSeek and Qwen now sweep the intelligence-per-dollar top tier, while 'Flash'/'mini'-branded closed models like Gemini 3.5 Flash cost over 20× more per Index run.