This is the monthly premium issue for paid subscribers, distilling the previous month of agentic coding updates into a single email.

You should be able to expense this under your team's learning and development budget. You can send this email or Slack message to your manager to ask for approval.

Welcome to the latest edition of ACW Monthly Brief. It's one email to catch you up on all the meaningful developments in agentic coding from July 2026.

In this issue, we will cover:

  1. Current state of agentic coding benchmarks

  2. Executive summary to catch up on the whole month in two minutes

  3. Important model releases from July, including GPT 5.6, Opus 5, and Kimi K3

  4. Tool updates related to Claude Code, MCP, and Grok Build CLI

  5. Trending open-source tools from July that I found useful

  6. 3 best agentic coding workflows from across the web

  7. Reading list for the month with 5 must-reads. Plus a longer list if you want to go deeper

  8. Miscellaneous updates and a few fun links from the agentic coding world

Non-AI disclaimer: This post is entirely my own work. Here’s why.

1. Coding Benchmarks

Current standing of open-weights and proprietary models on agentic coding benchmarks. The model names in bold were newly released in July.

Model

DeepSWE 1.1

Frontier Code 1.1 Main

Terminal-Bench 2.1

Pricing - per 1M I/O tok

Opus 5

74%

53.4%

-

$5 / $25

GPT 5.6 Sol

73%

47.5%

88.8%

$5 / $30

Fable 5 / Mythos 5

70%

53.5%

88.0%

$10 / $50

GPT 5.6 Terra

70%

41.3%

87.4%

$2 / $12

Kimi K3

69%

44.2%

88.3%

$3 / $15

Opus 4.8

59%

46.5%

66.1%

$5 / $25

Grok 4.5

54%

42.4%

83.3%

$2 / $6

DeepSeek V4 Flash 0731

54%

-

82.7%

$0.435 / $0.87

Muse Spark 1.1

53%

-

80.0%

$1.25 / $4.25

Gemini 3.6 Flash

49%

-

-

$1.5 / $7.5

GLM 5.2

44%

24.5%

81.0%

$1.4 / $4.4

DeepSeek V4 Pro

12%

17.6%

72.1%

$0.435 / $0.87

Inkling

-

14.0%

63.8%

$3.74 / $9.36

2. Executive Summary

  1. July was a busy month for agentic coding. We got GPT 5.6, Opus 5, Kimi K3, and several other models. Kimi K3 closed the gap between open-weights and proprietary models and is among the frontier models available today.

  2. GPT 5.6 family of models, Sol, Terra, and Luna are now publicly available. On coding benchmarks, Sol max seems to match or beat Fable 5 max performance.

  3. Kimi K3 is the largest open-weights models ever with a total of 2.8T parameters. It’s a SOTA model and sits among the 3 best models available today, Fable 5, GPT 5.6 Sol, and Kimi K3.

  4. Opus 5 is the latest Claude Opus model. Since Fable is only available on Max plans, it’s the best model available on all plans.

  5. DeepSeek released an update to the DeepSeek V4 Flash model. Beats the V4 Pro on all benchmarks and is at the pareto frontier for intelligence vs cost per task.

  6. Google release Gemini 3.6 Flash, XAI released Grok 4.5, and Meta released Muse Spark 1.1. Alibaba released Qwen 3.8 Max, 2.4T parameters model to compete with Fable 5 and Kimi K3. Qwen 3.8 Max weights are supposed to be released soon.

  7. Thinking Machines, the AI lab founded by Mira Murati, after leaving OpenAI, released their first model, Inkling, a 975B open-weights model.

  8. MCP received the biggest update since its release. It's now a stateless protocol instead of a stateful one.

  9. An unreleased OpenAI model breached Hugging Face internal network. Post-mortem of the incident has been published by the Hugging Face. They used GLM 5.2 for investigation cause closed models refused to help.

  10. OpenAI audited tasks in the SWE-Bench Pro and found that ~30% of the tasks are broken. No longer recommended to use this benchmark to assess agentic coding performance of models.

Before we dive into the entire month of updates, enjoy this banger from reddit:

3. Model Releases

There were quite a few releases in July including GPT 5.6, Opus 5, and Kimi K3. Looking at the pareto frontier of intelligence and cost per task, the most exciting one to me is the DeepSeek V4 Flash update. The Flash preview was already quite capable and this update has made the DSV4 Flash my default model for most things.

  1. Hy3 (Jul 6) - A new open-weights model from Tencent. 295B total parameters, 21B active. 256k context window. Supports only text inputs.

    1. Scores 57.9% on SWE-bench Pro and 71.7% on Terminal Bench 2.1. Agentic Coding benchmarks scores put it close to DeepSeek V4 Pro and Gemini 3.1 Pro which is pretty impressive for a model this size.

  2. GPT 5.6 (Jul 9) - GPT 5.6 family of models, Sol, Terra, and Luna, are now publicly available. On coding benchmarks, Sol max seems to match or beat Fable 5 max performance and on xhigh reasoning, seems to match/beat GPT 5.5 xhigh and beat Opus 4.8 xhigh.

    1. On Terminal Bench 2.1, Sol max scores 88.8% compared to 88.0% of Mythos 5/Fable 5. On xhigh reasoning though, which is what I use for coding tasks, GPT 5.6 Sol scores 84.9% on Terminal Bench 2.1 compared to 85.6% scored by GPT 5.5 at the same xhigh reasoning.

    2. Above the max reasoning level, there is a newer level, ultra, where the model will automatically coordinate among multiple agents to give the best answers.

    3. Sol is priced similarly to GPT 5.5, $5 / $30 per million input / output tokens. For comparison, Fable 5 costs $10 / $50 per million input / output tokens.

    4. On July 30, OpenAI reduced the API pricing for Terra by 20% and Luna by 80%. Current pricing is $2 / $12 per million input / output tokens for Terra, and $0.20 / $1.2 per million input / output tokens for Luna. Sol pricing is unchanged.

  3. Muse Spark 1.1 (Jul 9) - Latest model from Meta. Scores 61.5% on SWE-bench Pro and 80.0% on Terminal Bench 2.1, so it's behind Opus 4.8, Grok 4.5, and GLM 5.2. However, Pricing is $1.25 / $4.25 per million input / output tokens which is just less than GLM 5.2 but much cheaper than Grok 4.5.

  4. Inkling (Jul 15) - First model from Thinking Machines, the AI lab founded by Mira Murati, after leaving OpenAI. Inkling is open-weights, has MoE architecture with 975B total parameters and 41B active, 1M context window, and supports multimodal input (text, images, and audio).

    1. While the model has more parameters than GLM 5.2 (~753B / ~40B), the coding benchmarks are a bit underwhelming. Scores 54.3% on SWE Bench Pro and 63.8% on Terminal Bench 2.1. Both of these scores are much lower than GLM 5.2 which scores 62.1% and 82.7% on those benchmarks respectively. So this model is nowhere close to the frontier of the open-weight models.

    2. However, the USP seems to be the multimodal input, especially audio. It's the largest open weight model that supports audio input. And considering it's the first release from this AI lab, sounds like a good start. Let's see what they do in the future.

  5. Kimi K3 (Jul 16) - Latest kimi model with 2.8T total parameters and 104B active. It's the largest open-weight model ever.

    1. The coding benchmarks scores are only behind Fable 5 and GPT 5.6 and beat Opus 4.8 and GPT 5.5. It even beats both Fable 5 and GPT 5.6 on WebDev arena for front-end web development tasks.

    2. Pricing is $3 / $15 per million input / output tokens which is highest we've ever seen for an open-weights model. But it does make sense considering the size of the model and the performance.

  6. Grok 4.5 (Jul 16)- Cursor worked together with XAI on this new model. On SWE-bench Pro, it's performance is similar to Opus 4.7 and on Terminal Bench 2.1 it's similar to GPT 5.5. Context window is only 500k though.

    1. The differentiation seems to be speed and pricing. It runs at 80 TPS and generates 4x fewer token (on SWE-bench Pro tasks) than Opus 4.8 max.

    2. The pricing is $2 / $6 per million input / output tokens which seems cheap but in line with GLM 5.2 which performs similarly. However, there are two caveats with pricing:

      • $2 / $6 applies to context window less than 200k. Above that, pricing is twice, $4 / $12. In most coding I go over 200k easily so not it's not so cheap.

      • Cached tokens is priced at $0.5 per million tokens for context less than 200k and then $1. This is 25% of input token pricing. Typically, models (Opus, GPT, etc.) charge 10% of input token pricing for cached tokens and some (DeepSeek) even less. Considering cached tokens dominate agentic coding tasks, this model is not so cheap.

  7. Qwen 3.8 Max (Jul 19) - A 2.4T parameters model to compete with Fable 5 and Kimi K3. Not open-weights yet. Available only through the Alibaba Token Plan subscription.

  8. Gemini 3.6 Flash and 3.5 Flash-Lite (Jul 21) - 3.6 Flash is a bit more token efficient than 3.5 Flash. Output pricing is also a bit lower than 3.5 Flash, $7.5 per million output tokens vs $9. Still too expensive for what it's worth. 3.5 Flash-Lite improves upon 3.1 Flash-Lite and is served at 350 tokens/sec. More expensive though, $0.3 / $2.5 vs $0.25 / $1.5 for 3.1 Flash-Lite.

  9. Claude Opus 5 (Jul 24) - Newest Opus model. On DeepSWE and FrontierCode coding benchmarks, it comes close to Fable 5. On FrontierBench, successor to the Terminal Bench, Opus scores significantly better than the Fable 5 and GPT 5.6 Sol.

    1. Opus 5 was not trained on cyber tasks. It still comes close to Mythos 5 on finding cybersecurity vulnerabilities but remains significantly behind on exploiting those vulnerabilities.

    2. Pricing is same as the previous Opus models, $5 / $25 per million input / output tokens.

  10. DeepSeek-V4-Flash-0731 (Jul 31) - DSV4 Flash is out of preview with this update and beats even DSV4 Pro on all benchmarks.

logo

Upgrade to Paid Tier to continue reading.

Become a paying subscriber of Agentic Coding Weekly to get access to this post and other subscriber-only content.

Upgrade to Paid Tier

A paid subscription gets you:

  • Monthly brief on the 1st of each month
  • Weekly email every Monday

Reply

Avatar

or to participate

Keep Reading

© 2026 Agentic Coding Weekly.
beehiivPowered by beehiiv