Shipped in June: Private Mode, OAuth, GLM 5.2, Media Tools, and More

June was another very full month for NanoGPT.

Some of the work was highly visible: Private Mode became more complete, Sign in with NanoGPT launched, GLM 5.2 and Claude Sonnet 5 arrived, Batch API support expanded, the media catalog got a lot bigger, and payments gained more direct crypto and x402 options.

Some of the work was quieter but just as useful: clearer model pages, better payment progress, easier media waits, and many smaller fixes that make the product feel smoother day to day.

If you only skim one thing, the short version is:

• Private Mode became publicly available for eligible TEE models, with verification receipts and stronger local proxy support.
• Sign in with NanoGPT lets third-party apps use OAuth PKCE instead of asking users to paste API keys.
• The Batch API expanded beyond GPT, Claude, and Gemini to more open-model routes.
• GLM 5.2, GLM 5.2 Thinking, and GLM 5.2 TEE became available, including subscription access and private routes.
• Claude Sonnet 5 arrived as the latest Sonnet upgrade for coding, agents, long context, and knowledge work.
• Sofya, Linkup Research, Fugu Ultra, Kimi K2.7 Code, Qwen3.7 Plus, MiniMax M3, and more joined the model catalog.
• Image, video, and audio generation gained many new options, including Nano Banana 2 Lite, Qwen Image 2.0 Pro refresh, Seedance 2.0 Mini, Kling 3.0 Turbo, HappyHorse 1.1, Seed Audio, Stable Audio 3, and Fun-ASR Flash.
• Payments improved with Solana SOL, Solana USDC, Solana USDT, Lightning L402/x402, Solana USDC x402, THORChain re-enablement, and better balance flows.
• Everyday app workflows improved through smarter Global Memory, passkey-encrypted chat sync, post-history custom instructions, SVG previews, better attachments, and the new Cost Simulator.
• The community kept building, including CharacterVault, Nanoodle, Lumiverse, World Forge, OpenClaw, LettuceAI, and the Nano Noodle Cookoff.

For the full live changelog, see the Updates page.

Private Mode and privacy controls

The biggest privacy work in June was the public release and follow-through for Private Mode.

Private Mode is now publicly available for eligible TEE/* models in the NanoGPT web app. When enabled, the browser encrypts the private chat request before NanoGPT receives it. NanoGPT can still authenticate, route, and bill the request, but it cannot read the encrypted prompt or completion body.

Private Mode responses now include verification receipts. You can inspect the browser attestation result, TEE target, enclave host, verifier step statuses, measurement fingerprints, encryption flags, and receipt hashes. Developers can also use the local proxy:

NANOGPT_API_KEY=sk-your-key npx @nanogpt/private-mode

The package was updated again at the end of the month to support GLM 5.2 and GLM 5.2 Thinking through local OpenAI-compatible Private Mode clients, CLIs, agents, and other tools.

We also added and upgraded private model routes throughout the month, including Qwen3.6 27B TEE, Qwen3.6 35B TEE, DeepSeek V4 Flash TEE, GLM 5.2 TEE, and GLM 5.2 Thinking TEE.

The Privacy Guide now automatically scores more of your setup. It checks account access, payment history, model privacy, PII redaction, and storage sync settings, then shows which areas pass, partially pass, or need attention.

Sign in with NanoGPT and external apps

June also made NanoGPT easier to connect to external apps.

Sign in with NanoGPT lets third-party apps use OAuth PKCE so users can approve access in NanoGPT and receive a dedicated API key for that app. This is useful for local apps, coding agents, chat frontends, workflow builders, and other OpenAI-compatible clients that should not ask users to manually copy a long-lived API key.

Read more in Sign in with NanoGPT: OAuth PKCE for Local AI Apps.

Nanoodle is a good early example. It is a browser-only workflow builder for NanoGPT where users can build a workflow, run it with NanoGPT, then create, share, or export a standalone HTML app. Because it uses NanoGPT OAuth, it can connect users without running its own inference server or asking for manual API-key setup.

We also announced a Fluent partnership. Fluent, the native macOS AI assistant that works inside every Mac app, now has a built-in NanoGPT integration, balance tracking, request cost tracking, model card pricing, and an exclusive discount on NanoGPT text models.

API and developer platform

The Batch API kept getting stronger.

It launched for high-volume asynchronous /v1/chat/completions jobs, with JSONL uploads, batch creation, polling, cancellation, and output downloads. Batch jobs are 50% cheaper than normal synchronous chat completions for supported models, which makes them a better fit for large jobs that can wait.

Later in June, Batch API support expanded to more API-only open-model routes, including DeepSeek V4 Flash, DeepSeek V4 Pro, GLM 5.1, GLM 5.2, GPT OSS 120B, GPT OSS 20B, Kimi K2.5, Kimi K2.6, Kimi K2.7 Code, MiniMax M2.7, MiniMax M3, Nemotron 3 Ultra NVFP4, Qwen3.6 Plus, and Qwen3.7 Plus batch model IDs.

API keys also gained opt-in request and response logging for /v1/chat/completions. When enabled for a key, recent request and response bodies are stored temporarily and can be viewed from Usage row details. This is meant for debugging one key when you need it, not for turning logging on everywhere.

Inline moderation for API requests also landed. API clients can pass moderation: true or a specific moderation model on supported text, image, or video generation requests. If moderation flags the input, NanoGPT charges only the moderation check, skips generation, and returns content_policy_violation.

Other developer-facing improvements included API usage filtering by key, usage exports respecting the selected key, and clearer API request log access.

Models and model discovery

June brought a large wave of new text and reasoning models.

Some highlights:

• Claude Sonnet 5, Anthropic's newest Sonnet model for coding, agentic workflows, long-context reasoning, computer use, and practical knowledge work
• GLM 5.2, GLM 5.2 Thinking, GLM 5.2 TEE, and GLM 5.2 Thinking TEE
• Kimi K2.7 Code and Kimi K2.7 Code High-Speed
• MiniMax M3 and MiniMax M3 Thinking
• Qwen3.7 Plus and Qwen3.7 Plus Thinking
• Nvidia Nemotron 3 Ultra 550B and Thinking
• Nex N2 Pro
• Fugu Ultra
• Doubao Seed 2.1 Pro and Doubao Seed 2.1 Turbo
• Cohere North Mini Code
• Greg 2 Super and Greg 2 Ultra
• Linkup Research Low, Medium, High, and XHigh
• Sofya Research and Sofya as a web search provider

Model discovery improved too. Text model detail pages now show model weight links when available, and model pages gained more benchmark coverage, including Vectara hallucination data, FrontierCode, Best Value Benchmark Frontier, and more model comparison context.

Model cards also got clearer service tier badges, so you can see which models support options like Flex or Priority before starting a chat or making an API request. Model pages now also show average output tokens over the last 7 days, with percentile context for comparing verbose models against more efficient ones.

We also kept tuning routing and pricing: GLM 5.2 became cheaper and CrofAI pricing decreased.

Image generation

Image generation had another busy month.

New and improved image models included:

• Nano Banana 2 Lite, Google's fast lower-cost Gemini 3.1 Flash Lite Image model for drafts, edits, product mockups, UI layouts, posters, and infographics
• Qwen Image 2.0 Pro, refreshed to the latest June 2026 version, with the previous March version kept available
• Ideogram 4.0, especially useful for typography-heavy design work
• Reve 2 Create, Edit, and Remix
• MAI-Image-2.5, including image-to-image editing
• Krea 2 Medium Turbo
• Boogu Image and Boogu Image Edit
• SenseNova U1 Infographic
• Cosmos 3 Super Image

Image Templates also launched publicly in early June. They let you browse reusable image recipes with preview examples, pick a starting point, adjust the subject and creative details, then generate a prompt for the current or suggested image model.

Behind the scenes, we improved image model examples, prompt preservation across image mode switches, multi-model image layouts, image upload handling, LoRA asset durability, and gallery loading.

Video and audio

Video generation also expanded.

New or improved video models included:

• LTX-2.3 Quality, with text-to-video, image-to-video, audio-to-video, reference-video-to-video, HDR, LoRA inputs, aspect ratios, and synchronized audio
• LTX-2.3 Spicy Image-to-Video and LoRA Image-to-Video
• Kling 3.0 Turbo Standard and Kling 3.0 Turbo Pro
• HappyHorse 1.1
• Seedance 2.0 Mini, Seedance 2.0 Mini Spicy Image-to-Video, Seedance 2.0 Mini Video Extend, and Seedance 2.0 Mini Video Edit
• Wan 2.7 and Seedance 2.0 Spicy video models
• Pixelcut Video Background Remover
• Luma Ray 3.2 and UNI-1
• Cosmos 3 Super Image-to-Video

Audio and transcription also improved:

• ByteDance Seed Audio 1.0
• Seed Speech TTS 2.0
• Stable Audio 3
• Fun-ASR Flash
• ElevenLabs Scribe V2
• NVIDIA Nemotron ASR
• Gemini audio input
• Microsoft MAI-Transcribe 1.5 and MAI-Voice-2

Payments and account funding

June was a big month for payments.

Solana deposits now support native SOL, Solana USDC, and Solana USDT. We also added Solana-aware wallet checkout, deterministic Solana deposit wallets, Solana deposit reconciliation, and Solana USDC x402 payments.

Lightning L402/x402 payments launched, giving developers another accountless payment path for supported API calls. Accountless x402 API payments also gained public documentation and a machine-readable endpoint matrix so clients can inspect supported paid endpoints before using Nano exact or Base USDC exact payment flows.

THORChain payment options were re-enabled, including XRP, AVAX, and ATOM. Zcash deposits were briefly paused for maintenance after the NU6.2 network upgrade and then re-enabled with shielded and transparent address support.

Ethereum deposits also gained clearer progress tracking, so users can better see where a deposit is in the flow while waiting for credit.

The balance page and pay checkout also changed a lot. We redesigned account funding, added more balance methods to pay checkout, improved wallet and Nanswap options, added payment method search telemetry, made the mobile checkout flow clearer, and surfaced typical deposit credit times.

Chat, memory, files, and everyday workflow

A lot of June work focused on daily product feel.

We added a NanoGPT Help Chat model for quick product questions, plus image generation time estimates so media jobs are easier to understand while you wait.

Global Memory became smarter. Saved memories can now track freshness, expired memories can be cleaned up, and larger memory sets are ranked more selectively against the current request so unrelated details are less likely to clutter prompts.

Passkey-encrypted chat sync launched, giving users a stronger privacy option for synced chat storage. Preset and settings sync also improved, along with model preference sync and several startup and hydration fixes.

Conversations gained post-history custom instructions, which let you add a final reminder after chat history for format, tone, language, or other instructions that should remain fresh at the end of the prompt.

SVG code blocks now show previews directly in chat, with the option to download the preview as a PNG. Standalone markdown images render more cleanly, and more attachment types are supported, including FIT workout files, CMD, MODULE, BUILD, KML, VCF, Twig, Canvas, MTL, ShaderGraph, SFTP, SPEC, and other coding, configuration, project, map, contact, and design files.

We also added the Cost Simulator, which lets you pick a model you recently used and compare what the same recent requests would have cost on other models using current API pricing.

Community

The community kept shipping around NanoGPT in June.

CharacterVault added settings and tag-editor improvements, reasoning UI improvements, configurable sections, provider-list updates, lorebook tools, context-panel ordering, and Sign in with NanoGPT support.

World Forge grew into a richer roleplay world-building pipeline, with brainstorming and revision improvements, NPC memory work, and an Auditioner command for checking whether a character spec produces the expected behavior in a scene.

Lumiverse added NanoGPT PKCE sign-in and richer NanoGPT connection status details. OpenClaw's NanoGPT provider plugin received maintenance and dependency updates. LettuceAI 2.0 shipped with Companion Souls, relationship tracking, Director mode, branch-tree navigation, faster local models, audio input, inline scene images, and broader localization.

Nanoodle also launched the Nano Noodle Cookoff, a contest for building and sharing AI workflows or browser-exportable apps with Nanoodle and NanoGPT.

Milan also joined a THORChain podcast for a longer conversation about NanoGPT, crypto payments, privacy, models, and where AI access is going.

Summary

In June we made NanoGPT:

• more private, through Private Mode receipts, GLM 5.2 TEE, and better privacy scoring
• more connected, through Sign in with NanoGPT, Nanoodle, Fluent, and community PKCE integrations
• more programmable, through Batch API expansion, request/response logging, inline moderation, x402, and usage tools
• stronger for coding and agents, through GLM 5.2, Claude Sonnet 5, Kimi K2.7 Code, Fugu Ultra, MiniMax M3, Qwen3.7 Plus, and more
• better for creative work, through new image, video, audio, transcription, LoRA, and template workflows
• easier to fund, through Solana, THORChain, Lightning/x402, better checkout flows, and clearer deposit progress

That is the direction we want to keep going: many models and tools, but with better privacy, clearer product context, stronger developer primitives, and fewer sharp edges in everyday use.

Thanks for using NanoGPT. We'll keep improving.