Source briefing. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Curious readers and builders getting oriented
⚡ TL;DR — the 30-second version
- Current reviewed lineup: Opus 5.5, GPT-6 Astra/Sol/Luna, Gemini 3.8 Flash and Grok 4.7.
- Prices are conditional: check long-context, caching, speed tiers and introductory expiry.
- Voice, image and cyber variants have different interfaces and access rules.
- Published provider scores are not independent DARC tests.
Anthropic, OpenAI, Google and xAI have changed the practical model shortlist within weeks. This 23 September 2026 field guide connects the releases to the decisions a builder actually faces: model IDs, output modalities, tool support, cost boundaries and deployment access. Follow the source links for the exact terms behind each entry.
Anthropic released Claude Opus 5.5 on 22 September. Its model ID is claude-opus-5-5, default context is 1M, and standard pricing is $4 input and $20 output per million tokens. Cache reads are $0.20/M. Thinking is always on. Anthropic's reported 40% reduction in typical workload cost is not the same as the 20% reduction in input/output token prices. Sonnet 5.5 and Haiku 5.5 are described as forthcoming, so this site does not mark them available.
“A current model map needs more than fresh names: exact API IDs, access rules, pricing conditions and evidence dates belong next to every recommendation.”
OpenAI introduced GPT-6 Astra on 3 September and added GPT-6 Sol and GPT-6 Luna on 22 September. The current API pages list $10/$50, $2/$10 and $0.10/$0.50 standard input/output prices per million tokens. All three have a 1.05M total context window; the headline prices apply to prompts up to 272K input tokens, after which long-context multipliers apply. Tool support and permitted reasoning parameters differ by endpoint and model.
Google's current general-purpose Flash model is Gemini 3.8 Flash, released 2 September. It accepts text, images, video, audio and PDF input and generates text. Its $0.75/$3.75 introductory token rates run through 31 December 2026. Gemini 3.8 Live and Live Extended Thinking are separate audio models, released 15 September. On 23 September Google added Gemini 3.8 Flash TTS and Flash-Lite TTS. Flash Cyber has restricted access through Fairwind; it is not interchangeable with the public Flash endpoint.
Grok 4.7 arrived on 21 September. xAI lists 500K context and $2 input / $6 output per million tokens, with a fast variant at twice the price. The launch table reports 46.3% on CursorBench 4.0 and 37.6% on Terminal-Bench 4.0 at xhigh effort. Those are xAI's published evaluations, not measurements performed for this website.
Open-weight options need the same care. DeepSeek's latest reviewed release is V4.1 Flash, served as deepseek-flash; the corrected changelog says V4 Pro remains available. Qwen3.8-Flash-Next is an open checkpoint with 262K native context, not the same deployment as the hosted 1M-context Qwen3.8-Flash. Mistral Medium 3.5 is an active coding model under a Modified MIT license—not a newly released September model. Use the model index for the individual sources and caveats.
Field guide / 01
Build a shortlist from constraints, not a leaderboard
Start by writing the non-negotiables: accepted input and output modalities, allowed processing regions, data-retention requirements, tool support and the budget for a completed task. Eliminate candidates that fail these constraints before comparing a benchmark.
Then select a small representative evaluation set. Include at least one task with missing evidence and one that must refuse an unauthorized action. A model that wins a public coding test can still be the wrong choice for a document workflow requiring reliable abstention.
| Decision | Check in documentation | Check in your application |
|---|---|---|
| Capability | Exact model and endpoint | Accepted-task rate |
| Cost | Cache, output and long-context rates | Retries and human review |
| Availability | Plan, region and access class | Rate-limit and failure handling |
| Safety | System card and safeguards | Tool and data boundaries |
Field guide / 02
An upgrade is a controlled experiment
Run the old and proposed model against the same frozen inputs without allowing both to perform real external actions. Compare results blind where practical, retain failures and record total usage. A shadow evaluation should not create duplicate customer messages or records.
Roll out only after critical cases pass, then keep a rollback path and usage alerts. The recommendation is deliberately conditional: the current model map tells you what to test, not which provider must always win.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Should every app switch immediately to the newest model?
No. A release is a reason to reevaluate. Switch when your own tests show useful gains without breaking cost, privacy, reliability or integration requirements.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Introducing Claude Opus 5.5Anthropic↗
- 02Claude models overviewAnthropic↗
- 03OpenAI API changelogOpenAI↗
- 04GPT-6 Astra API model and pricingOpenAI↗
- 05GPT-6 Sol API model and pricingOpenAI↗
- 06GPT-6 Luna API model and pricingOpenAI↗
- 07Gemini 3.8 Flash — introductory pricing and migrationGoogle↗
- 08Introducing Grok 4.7xAI / SpaceXAI↗
- 09DeepSeek changelog — V4.1 Flash and V4 Pro correctionDeepSeek↗
- 10Qwen3.8-Flash-Next official model cardQwen↗
- 11Mistral Medium 3.5 model cardMistral AI↗
