Pulse

Research preview for designs, prototypes, slides, decks. Derives brand systems from your codebase. Free for Pro/Max/Team/Enterprise.

Reasoning model for biology, drug discovery, translational medicine. Bundled with Codex research plugin connecting to 50+ scientific tools.

DeepMind release improving spatial pointing, counting, success detection, and instrument reading. Available on Gemini API + AI Studio. Prior 1.5 sunsets April 30.

SWE-bench Verified 87.6% (from 80.8%), CursorBench 70%, GPQA Diamond 94.2%. Adds xhigh effort, Task budgets, /ultrareview multi-agent review. Same $5/$25 pricing as 4.6.

30B activated MoE targeting agentic coding. Apache-style open weights on Hugging Face. Leads Gemma 4 on several agentic coding benchmarks.

Orchestrates multi-step workflows across Photoshop, Premiere, Lightroom, Express, Illustrator. Public beta incoming.

GPT-5.4 fine-tuned for cyber defense including binary reverse engineering. Access gated via expanded Trusted Access for Cyber program.

HD mode 3x faster and 3x cheaper; standard res 50% faster / 25% cheaper. Image prompts and image weights now supported. New Prompt Shortener tool.

Both models scheduled for retirement on June 15, 2026 on the Claude Developer Platform. Migrate to 4.5 or 4.7.

229B sparse MoE (10B active) on Hugging Face / ModelScope under Modified-MIT. 56.22% on SWE-Pro, 57.0% on Terminal Bench 2.

New fallback model replacing GPT-5 Instant Mini after rate limits. Smoother writing and context handling on the cheaper tier.

Text/image/video/audio inputs with 15s multi-shot output, native stereo audio, director-level camera control.

Native multimodal reasoning, tool use, visual chain-of-thought, multi-agent orchestration. Available in Meta AI app, meta.ai, private API preview. Ground-up overhaul from Llama.

Agent loop, tool execution, sandbox, state persistence all via REST. $0.08/session-hour + API token costs. Header-gated: managed-agents-2026-04-01.

Anthropic research preview positioned as new model class with SOTA cyber/code/reasoning. Limited to internet-critical orgs + OSS maintainers. On Bedrock and Vertex AI.

40B active, 200K context. 58.4 on SWE-Bench Pro (edges GPT-5.4, Claude Opus 4.6). MIT license on HF. API: $1/$3.20 per M tokens.

Use text, image, video, or audio to generate or edit videos. Available on Unlimited/Enterprise plans outside the US.

First/last-frame control, video-to-video editing, subject referencing. Launched on Together AI April 3.

Business/Enterprise/Edu Custom GPTs lost GPT-4o access (after Feb 13 consumer retirement). API access still available.

$0.05/sec at 720p, $0.08/sec at 1080p. Paid preview on Gemini API, AI Studio, Vertex AI.

New texturing tools for faster, smarter 3D content creation on existing models.

Up to 7 songs/day with commercial-safe licensing (Merlin + Kobalt deals). Competes with Suno and Udio.

1M-token context window. Proprietary flagship for enterprise agentic AI deployments.

4 sizes (E2B, E4B, 26B-A4B MoE, 31B dense). 140+ languages, function calling, audio/video inputs. 31B ranks #3 on Arena open leaderboard.

Low-poly asset generation in ~2 seconds. Tripo AI frames it as the start of 'AI 3D 2.0.' Debuted at GDC 2026.

Thinker-Talker + Hybrid-Attention MoE. 256K context, 113 languages. SOTA on 215 audio/video benchmarks. Kept closed-source.

9 languages, 5-sec voice cloning, TTFA 90 ms, RTF 6x. Human-preference beat ElevenLabs Flash v2.5 68.4% of the time.

128K context, 90+ languages. Powers upgraded Gemini Live and Search Live in 200+ countries.

Upgrade from Lyria 3 (30s cap): 3-min tracks with structural control over intros/verses/choruses/bridges. In Gemini app (paid), Google Vids, Vertex AI preview.

Web/app end April 26, API end September 24. OpenAI cites compute cost (~$1M/day dev, ~$15M/day inference vs $2.1M lifetime revenue). Disney walks from $1B deal.

262K context, $0.40/$2.00 per M tokens on OpenRouter. Tops BigBench Audio (94.0), MMAU-Pro (69.4).

Anthropic's most capable model now supports 1 million token context. Pricing unchanged. Available via API and platform immediately.

Vidu deprecated bare "viduq2" model name. Must now use "viduq2-pro", "viduq2-pro-fast", "viduq3-pro", "viduq3-turbo", "viduq1", or "vidu2.0". No announcement from Vidu — discovered via our API monitor when requests started failing.

Google added image generation to Gemini 3.1 Flash via the standard generateContent endpoint. Supports 4K output, web search grounding, and character consistency. $0.08/image at 1K resolution.

Tested Bria's API: background removal, 2x/4x upscaling, image expansion, background replacement — all synchronous, instant results. No cold starts, no queues. $0.018-$0.04/image. This is what fal.ai should have been.

Kling AI added 2.6 Pro and Standard variants via Together.ai. $0.14/sec. Early testing shows significant motion improvement over 2.5 Turbo, especially for camera movements and character animation.

Both models return "unable to access model" errors. Removed from our platform. No deprecation notice from Together or Kling — just stopped working.

After extensive testing: Veed.io's API redirects to fal.ai. HuggingFace Inference Providers routes to fal.ai, Replicate, or WaveSpeed. fal.ai itself is serverless GPU with 10+ minute cold starts on less popular models (Beatoven, Cosmos). Direct APIs are faster and cheaper.

First 3D model generation service integrated. Text or image → GLB model with PBR textures in ~60 seconds. Interactive viewer with auto-rotate. $0.20/model, 3 modes (text, image, multiview).

New Qwen3.5 Plus model from Alibaba Cloud. Competitive pricing at $0.0008/1K input, $0.002/1K output. 128K context window. Strong multilingual performance, especially CJK languages.

Chatterbox model now supports multiple languages beyond English. New chatterbox-multilingual variant available. Voice cloning quality remains best-in-class for conversational TTS.

New: Claude Opus 4.6, Gemini 3.1 Flash Image, Kling 2.6 Pro/Standard, Qwen3.5 Plus, Chatterbox Multilingual, FLUX 2 Klein 4B/9B, Tripo V2.5 (3 modes). Deprecated: Kling O1 Pro/Standard, viduq2 bare name. Breaking: Vidu model names, Bria enhance sync mode removed for 4MP+ images.