August's flagship model releases: GPT-5, Opus 4.1, gpt-oss and DeepSeek V3.1 compared

August was unusually dense with model releases. Within three weeks, Anthropic shipped Claude Opus 4.1, OpenAI shipped both GPT-5 and its first open-weight models since GPT-2, and DeepSeek released a hybrid reasoning model under an MIT license. For a buyer, the useful question is not which one "won" but what each release changes about cost, deployment options and risk.

The releases

Claude Opus 4.1 (August 5)

Anthropic positions Opus 4.1 as an upgrade to Opus 4 for agentic tasks, real-world coding and reasoning, and reports 74.5% on SWE-bench Verified, measured without extended thinking ^[1]. Pricing is described only as "the same as Opus 4" ^[1]. It is available to paid Claude users, in Claude Code, via the API, Amazon Bedrock and Google Cloud's Vertex AI ^[1]. Anthropic also said it planned "substantially larger improvements" in the coming weeks ^[1].

GPT-5 (August 7)

OpenAI describes GPT-5 as a unified system that routes between a faster model and a deeper reasoning model depending on the query; the secondary coverage notes OpenAI has not confirmed whether this is one model or several orchestrated ones ^[2]. Reported API pricing is $1.25 per million input tokens and $10.00 per million output tokens, with GPT-5 mini and nano priced lower and cached input at one-tenth of the live input price ^[2]. Reported SWE-bench Verified: 74.9%, versus 69.1% for o3 ^[2].

OpenAI's GPT-5 system card reports that on prompts representative of ChatGPT production conversations, gpt-5-main produced 44% fewer responses with at least one major factual error than GPT-4o, and gpt-5-thinking 78% fewer than o3, as scored by an LLM grader with web access that agreed with human reviewers 75% of the time ^[5]. It also reports that a chain-of-thought monitor flagged deception in about 4.8% of o3 responses and about 2.1% of gpt-5-thinking responses on a representative set of conversations ^[5]. These are first-party figures, each against a different baseline, and the system card notes that the 74.9% SWE-bench result comes from the launch post and was run at default API verbosity ^[5].

gpt-oss (August 5)

OpenAI released two open-weight models under the Apache 2.0 license: the 120B model (117 billion parameters, 5.1 billion active per token) and the 20B (21 billion, 3.6 billion active) ^[3]. The larger runs on a single 80 GB H100; the smaller fits in 16 GB of VRAM ^[3]. Context is 128K tokens, reasoning effort is set to low, medium or high in the system prompt, and the models lack vision ^[3]. OpenAI says the 120B performs roughly on par with its proprietary o4-mini and the 20B is similar to o3-mini ^[3].

DeepSeek V3.1 (August 21)

DeepSeek V3.1 has 671 billion total parameters with 37 billion active, a 128,000-token window in both modes, and switches between a "think" mode and a "non-think" mode ^[4]. Code and weights are MIT licensed ^[4]. Per The Decoder, Artificial Analysis scored it 60 on its intelligence index in reasoning mode versus 59 for R1, and noted it does not support function calling in reasoning mode, a limit for agent workflows ^[4].

Side by side

DeepSeek's API pricing, starting September 5, is $0.07 per million input tokens on a cache hit, $0.56 on a cache miss and $1.68 for output ^[4].

How to read this

The gap between 74.5% and 74.9% is smaller than the variance you should expect from harness choices, tool setups and run counts. A critic quoted in coverage of GPT-5 made a related point, noting the gain over the previous state of the art was only 0.4% ^[2]. What differentiates these releases for a business is less the benchmark and more the operating model: closed models with usage pricing, or open weights you can host yourself under permissive licenses ^[3][4].

The open-weight options matter for data-handling constraints. A 20B model that fits in 16 GB can run on a single workstation, and a 120B model on one 80 GB accelerator ^[3]. Neither claims frontier-level capability; OpenAI's own comparison is to o4-mini and o3-mini ^[3].

Practical implication

Treat public benchmark gaps of a point or less as ties, and run your own task-level evaluation on a few hundred of your real cases before switching providers. If data residency or per-token cost is the constraint, the August open-weight releases make self-hosting a credible option to price out.