benchr — AI model reviews and field guides https://benchr.org/ 2026-07-29T00:00:00Z The benchr team Claude Opus 5: the migration question behind Anthropic's new flagship https://benchr.org/articles/claude-opus-5-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z The same $5/$25 list price, a 1M-token window, and one configuration change that can break a careless upgrade. Gemini 3.1 Flash TTS Preview: steerable speech with preview boundaries https://benchr.org/articles/gemini-3-1-flash-tts-preview-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Google's text-to-speech model has explicit token limits and Batch support, but no public per-token price in the checked model card. Gemini 3.5 Flash-Lite: the $0.30 subagent target has a migration cost https://benchr.org/articles/gemini-3-5-flash-lite-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Google's lowest-cost 3.5 model is built for throughput, but the API changes are part of the selection decision. Gemma 4 26B A4B IT: the MoE choice that still needs real VRAM https://benchr.org/articles/gemma-4-26b-a4b-it-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Only 3.8B parameters activate per token, but all 25.2B weights stay loaded. Here is the deployment decision behind Gemma 4 26B A4B IT. Gemma 4 31B IT: the dense quality-first choice https://benchr.org/articles/gemma-4-31b-it-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Gemma 4 31B IT review: dense architecture, official memory estimates, 256K context, sibling results, deployment fit, and when to avoid it. GPT-Live-1 is a ChatGPT Voice rollout, not an API endpoint https://benchr.org/articles/gpt-live-1-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z OpenAI's full-duplex Voice model changes the product experience while leaving a developer pricing checklist empty. GPT-Live-1 mini: the Voice fallback is not a developer SKU https://benchr.org/articles/gpt-live-1-mini-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z OpenAI's smaller full-duplex ChatGPT Voice model reaches a different user tier, with no public API contract yet. GPT-Realtime-2.1 mini: the lower-cost voice model still has an audio bill https://benchr.org/articles/gpt-realtime-2-1-mini-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z OpenAI's distilled realtime reasoning tier keeps the large context window while changing the cost shape. GPT-Realtime-2.1: evaluate the conversation loop, not a text-only price card https://benchr.org/articles/gpt-realtime-2-1-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z OpenAI's speech-to-speech model adds reasoning and tool use, but its audio bill needs its own test plan. Leanstral 1.5 is for proofs, not chat: read its benchmark claims in context https://benchr.org/articles/leanstral-1-5-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Mistral's open-weight formal-verification specialist has an unusual architecture and provider-reported proof metrics—not a general chatbot score. Muse Image: Meta's consumer launch leaves a developer checklist blank https://benchr.org/articles/muse-image-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z The image model is available in Meta AI, but the checked announcement gives developers no public model ID, rate card, or token limits. Muse Video Preview: native audio is a preview feature, not an API contract https://benchr.org/articles/muse-video-preview-launch 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Meta previewed a video model with native audio alongside Muse Image, while leaving public developer terms unannounced. Qwen-Audio 3.0 TTS Flash: latency and cloning change the deployment decision https://benchr.org/articles/qwen-audio-3-0-tts-flash-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Alibaba's low-latency TTS branch adds voice cloning, so consent and call-flow tests matter alongside speed. Qwen-Audio 3.0 TTS Plus: built-in voices are the product boundary https://benchr.org/articles/qwen-audio-3-0-tts-plus-review 2026-07-28T00:00:00Z 2026-07-29T00:00:00Z Alibaba's instruction-controlled TTS model targets standard synthesis, with no public token limits or per-token price in the checked docs. Gemini 3.6 Flash launch: cheaper output, same Flash input https://benchr.org/articles/gemini-3-6-flash-launch 2026-07-22T00:00:00Z 2026-07-29T00:00:00Z Google released Gemini 3.6 Flash on July 21, 2026. It keeps Gemini 3.5 Flash's $1.50 input price, cuts output to $7.50, and becomes Google's recommended Flash… Correction: Gemini 3.5 Pro has not been released https://benchr.org/articles/gemini-3-5-pro-review 2026-07-01T00:00:00Z 2026-07-29T00:00:00Z benchr previously published a review and pricing page for "Gemini 3.5 Pro." That model has not launched. Here is the correction, and what to read instead. Claude Sonnet 5: Mythos-class architecture at Sonnet pricing https://benchr.org/articles/claude-sonnet-5-launch 2026-07-01T00:00:00Z 2026-07-23T00:00:00Z Claude Sonnet 5 launched July 1, 2026 with introductory $2/$10 pricing through August 31, then $3/$15 standard pricing from September 1. GPT-5.6 Sol, Terra, and Luna: OpenAI's frontier series is generally available https://benchr.org/articles/gpt-5-6-launch 2026-06-28T00:00:00Z 2026-07-13T00:00:00Z GPT-5.6 (Sol, Terra, Luna) is generally available via the OpenAI API as of July 9, 2026. Pricing, benchmarks, context, and specs verified against official… How to fine-tune an open model without owning a GPU https://benchr.org/articles/fine-tune-open-model-no-gpu 2026-06-19T00:00:00Z 2026-07-23T00:00:00Z A practical QLoRA guide for fine-tuning 7B–70B open models on rented GPUs, with clearly labelled June 2026 planning scenarios. “CUDA out of memory”: why it happens and how to fix it https://benchr.org/articles/cuda-out-of-memory 2026-06-19T00:00:00Z 2026-06-19T00:00:00Z The five real causes of the CUDA out of memory error when you serve an LLM, the fixes in order from free to last resort, and when you just need a bigger GPU. Your computer can't run the big open models. Here's what actually works. https://benchr.org/articles/run-big-models-without-a-gpu 2026-06-19T00:00:00Z 2026-06-19T00:00:00Z Why DeepSeek and Llama 70B won Renting a GPU vs. paying per token: when self-hosting an open model is actually cheaper https://benchr.org/articles/self-host-vs-api-gpu-cost 2026-06-19T00:00:00Z 2026-06-19T00:00:00Z The honest break-even math for self-hosting an open model on a rented GPU versus a per-token API, and why utilization, not the sticker price, decides it. The U.S. Pulled Claude Fable 5 and Mythos 5: What It Means https://benchr.org/articles/fable-mythos-export-control 2026-06-16T00:00:00Z 2026-07-23T00:00:00Z The complete June 2026 Fable 5 and Mythos 5 export-control timeline: suspended June 12, controls lifted June 30, access restored July 1, and what developers… Claude Fable 5 vs GPT-5.5 vs Gemini 3.1 Pro https://benchr.org/articles/fable-5-vs-gpt-5-5-vs-gemini-3-1-pro 2026-06-13T00:00:00Z 2026-07-23T00:00:00Z Claude Fable 5 vs GPT-5.5 vs Gemini 3.1 Pro: the three newest frontier models compared on price, context, published benchmarks, and the right model per use… Which Claude model should you use in 2026? https://benchr.org/articles/which-claude-model 2026-06-13T00:00:00Z 2026-07-23T00:00:00Z A decision guide to Anthropic Claude Opus 4.8 vs Gemini 3.1 Pro https://benchr.org/articles/opus-4-8-vs-gemini-3-1-pro 2026-06-13T00:00:00Z 2026-06-13T00:00:00Z Claude Opus 4.8 vs Gemini 3.1 Pro: Opus leads coding and agentic work, Gemini is ~2.5x cheaper on input and edges abstract reasoning. GPT-5.4, reviewed: the value pick OpenAI doesn't advertise https://benchr.org/articles/gpt-5-4-review 2026-06-10T00:00:00Z 2026-07-29T00:00:00Z GPT-5.4 review, three months after launch: $2.50/$15 pricing, 1M context, 75% OSWorld computer use, and why losing the flagship crown to GPT-5.5 made it the… Claude Fable 5: the Mythos-class model you can finally use https://benchr.org/articles/claude-fable-5-launch 2026-06-10T00:00:00Z 2026-07-23T00:00:00Z Anthropic restored Claude Fable 5 access after its June suspension. The model costs $10/$50 per million tokens. Claude Opus 4.8 Is Live: Coding, Agents, and Mythos Pressure https://benchr.org/articles/claude-opus-4-8-launch 2026-06-09T00:00:00Z 2026-07-23T00:00:00Z Claude Opus 4.8 shipped May 28, 2026 at the same $5/$25 pricing as 4.7. The real story is agentic coding, dynamic workflows, and a Mythos-class model heading… Anthropic Expands Project Glasswing and Ships Claude Security https://benchr.org/articles/anthropic-glasswing-claude-security 2026-06-09T00:00:00Z 2026-06-10T00:00:00Z Anthropic extended Project Glasswing to about 150 more organizations and launched Claude Security, a product that uses Claude Opus 4.8 to scan codebases and… Google I/O 2026: Developer AI Moves From Autocomplete to Agents https://benchr.org/articles/google-ai-agents-io-2026 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z At I/O 2026 Google shipped Gemini 3.5 Flash as an agent-first model, started moving Gemini CLI users to Antigravity by June 18, and reframed developer AI… Google's AI Search Is Becoming an Agent Layer, Not a Summary Box https://benchr.org/articles/google-ai-search-agent-layer 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z Google's AI Mode is shifting from summarizing results to acting on them with background agents. Meanwhile AI Overviews keep cutting clicks. GPT-5.5 Pricing Explained: The 272K Cliff and the $30/$180 Pro Tier https://benchr.org/articles/gpt-5-5-pricing-explained 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z GPT-5.5 costs $5/$30 per million tokens — until a prompt crosses 272K input tokens, when the whole session reprices higher. Grok 4.3 Is Now the Default — and Old API Slugs Bill at Its Prices https://benchr.org/articles/grok-4-3-default-and-pricing 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z xAI made Grok 4.3 the default model and redirected deprecated text slugs to it — billed at Grok 4.3 prices. The UK Just Forced Google to Give Publishers Control Over AI Search https://benchr.org/articles/uk-google-ai-search-publishers 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z The UK's CMA now requires Google to let publishers opt out of having their content power AI features in Search — without losing standard rankings. WebMCP Is a Warning: Websites Need to Become Agent-Readable Tools https://benchr.org/articles/webmcp-agent-readable-websites 2026-06-09T00:00:00Z 2026-06-09T00:00:00Z WebMCP is a proposed web standard, backed by Google and Microsoft, that lets sites expose structured tools to AI agents instead of making them guess from the… Lowest-Cost Hosted LLM APIs Tracked by benchr: July 24, 2026 Snapshot https://benchr.org/articles/cheapest-llm-api-2026 2026-06-06T00:00:00Z 2026-07-24T00:00:00Z A July 24, 2026 snapshot of the lowest-cost hosted LLM APIs tracked by benchr, including DeepSeek V4-Flash, GPT-5 Mini, and Claude Haiku. DeepSeek vs OpenAI Pricing: Cost Comparison & Quality Trade-offs https://benchr.org/articles/deepseek-vs-openai-pricing 2026-06-06T00:00:00Z 2026-07-24T00:00:00Z Is DeepSeek really 90% cheaper than OpenAI? A head-to-head pricing comparison of DeepSeek V4-Pro/Flash vs GPT-5.5/5. Sourced from official docs. OpenAI API Pricing Guide: GPT-5.5, GPT-5, and GPT-5 Mini Costs https://benchr.org/articles/openai-api-pricing-guide 2026-06-06T00:00:00Z 2026-07-24T00:00:00Z Dated OpenAI API pricing snapshot for listed GPT models, with input, output, caching, eligible batch rates, and official source links. AI Model Pricing Comparison 2026: Cost per Million Tokens https://benchr.org/articles/ai-model-pricing-comparison 2026-06-06T00:00:00Z 2026-06-06T00:00:00Z Compare API token pricing (input, output, caching) across OpenAI, Anthropic Claude, Google Gemini, DeepSeek, and open-weights. Verified pricing table.