benchr
Worth reading
DeepSeek-V4.1-Flash
DeepSeek · Replacing the retired DeepSeek V4-Flash on the same API, at a lower listed rate
See what AI can really do — and how people are doing it
What people built this week, how to do it yourself, what to use, and what to do when it breaks. Every claim names its source and its date.
Someone did this
57 working web demos, built by agent fleets in one day
The review step was handled by the agents too, which is the part that usually does not survive contact with a deadline.
Reported by elder-pliniusSeptember 1, 2026Not tested by benchr
What people are making
All discoveries →- Given a broken setup, the model installed its own toolingRepairing its own working environment was not the task it was given; it was what the task required first.Reported by The Neuron (Grant and Corey)September 2, 2026Not tested by benchr
- From one photo of an empty lot to a filmed tour of a houseThe output is a finished video, and the model got there by writing code rather than by using a modelling application.Reported by @alexalbert__September 1, 2026Not tested by benchr
- An agent you talk to on WhatsApp is now one of the most-starred projects on GitHubThe interface is the surprise. No terminal, no dashboard, no browser tab — the agent is a contact in a chat app you already have open.Reported by Peter SteinbergerAugust 30, 2026Not tested by benchr
Things you can do with AI
Start from a goal →Try this
All techniques →- Put the rule lastA long reference plus your rules → The rules stop losing to the document
- Make it restate the briefA brief with constraints that matter → Misreadings surface before they cost you a draft
- Ask for the last page firstA long document you just uploaded → You know whether the whole file arrived
Something not working?
All fixes →What providers changed
All changes →Explore the record
Open Explore →Which model, what it can do, what it costs, and what changed. Every figure names the page it was read from and the date.
One clear model decision
Compare → test → choose.
Three steps. No universal winner.
01
CompareRead the recordPrice · context · sources
02
TestUse your own casesBlind scores · real outputs
03
ChooseSave the shortlistFit · cost · evidence
38verified models
25logged changes
41retirement records
4 labelsevidence categories
New and changed
Recent model updates
LoadingReading the verified change record…
Latest
View every release →-
01
DeepSeek-V4.1-Flash
DeepSeek · Replacing the retired DeepSeek V4-Flash on the same API, at a lower listed rate
-
02
GPT-6 Astra
OpenAI · Long end-to-end work OpenAI aims this model at - reasoning, coding, computer use, research
-
03
Gemini 3.8 Flash: same rate card, different bill
Gemini 3.8 Flash is priced to the cent like 3.7 Flash. Google says it spends more tokens on purpose, which is where the cost difference lives.
-
04
Claude Fable 5.1: cheaper cache reads, stricter contract
Claude Fable 5.1 keeps the $10/$50 API rate and 1M context, cuts cache reads to $0.25 per 1M tokens, and changes forced-tool and thinking behavior.
-
05
Qwen3.8-Flash
Alibaba (Qwen) · High-volume multimodal work - among the lowest published rates benchr records
-
06
GLM-5.3: 1M coding, hosted now with weights pending
Z.AI's API release adds post-training gains, always-on reasoning, and three compatible protocols at $1.40/$4.40.
-
07
Gemini 3.7 Flash: introductory pricing and migration
Google's stable multimodal model starts at $0.75/$3.75 through 2026, with a 1M-token window and a dated price change.
-
08
Grok 4.6: 500K agent model with no numeric output ceiling
Grok 4.6 targets long coding agents, starts at $2/$6, doubles at 200K prompt tokens, and has no numeric text-output cap.
Reviews
-
07
Claude Opus 5
Anthropic's current Opus at the same $5/$25 rate — what changes over 4.8, and who should move.
-
08
GPT-5
A candidate for visual and structured-output work based on OpenAI's published positioning; validate it on your own prompts.
-
09
Gemini 3 Pro
A retired multimodal model that Google deprecated in March 2026 in favor of Gemini 3.1 Pro.
Comparisons
-
10
Coding assistants
Cursor, GitHub Copilot, Windsurf, and Cody on the same Markdown-exporter task.
-
11
Voice models
ElevenLabs, OpenAI Whisper, and Cartesia Sonic on latency, accuracy, and naturalness.
-
12
Context windows
Advertised context capacity versus reproducible retrieval checks you can run on your own documents.
Analysis
-
13
Price per use case
The cheapest model for chat, coding, RAG, agents, classification, and summarization.
-
14
The open-weight tier
Where Llama 4, Mistral Large 3, DeepSeek-V4, and Qwen 3.6 stand against the closed labs.
-
15
Why benchmarks stopped telling you
MMLU is saturated above 90%. The benchmarks worth tracking now.
Guides
-
16
Frontier models
Claude Opus 5 and Fable 5.1, GPT-5.6 Sol and GPT-6 Astra, Gemini 3.1 Pro and Grok 4.6: sourced specifications, editorial tradeoffs, and a local evaluation plan.
-
17
Open-weight models
Llama 4, Mistral, DeepSeek, Qwen, the small-model tier, and what it takes to self-host.
-
18
AI costs
What AI costs by model and workload, and where teams overspend.
Compare models directly
An interactive comparison covering pricing, benchmarks, context windows, and capability ratings for all 38 models in the shared index, across frontier, mid, and open-weight tiers.
All 38 models ranked by benchr Rating →
All 114 articles in the archive →
Provider guides: OpenAI, Anthropic, Google, and open weights →
benchr is an evidence-led reference, not a generic listicle. Figures are identified as official provider data, third-party benchmark results, or benchr editorial estimates so readers can judge the evidence behind each claim.
- Traceable evidence. Official prices and specifications link to provider documentation; benchmark results link to their maintainers; estimates and analysis are labeled as such. Verification dates and source fields are published in our open data file.
- A record no one else keeps. An append-only price history and a permanent deprecation register; retired models are never deleted — that memory is the point.
- Editorially independent and commercially transparent. Some marked links to AI/ML API, Lovable, and RunPod can earn benchr a commission without changing your price. Partners do not buy rankings or approve coverage. Google AdSense may be added only after approval. Read the affiliate disclosure, editorial standards, and corrections policy.