State of AI Coding Tools, Q3 2026: A Data Roundup

Research-based — not hands-on tested. Every figure on this page is a vendor-published figure or a third-party benchmark result, cited to a dataset row. We did not run any of them. How we review.

The quarter at a glance

Six findings, each with the rows that support it further down the page.

1. What shipped in Q3 2026

The table lists every model in the dataset with a sourced release date between 2026-07-01 and 2026-09-30. Prices are the rows in force on 2026-09-30. Where a promotional row is active it is shown instead of the standard row, the same rule the explorer uses. Prices are US dollars per million input and output tokens.

† The GPT-5.6 release date comes from a secondary source in the dataset: the OpenAI Developer Community quoting OpenAI’s launch post. Qwen3.8-Max is left out of the table because the dataset holds no source for its release date.

Two things stand out in the price column. First, the newest rows (Opus 5.5, Sonnet 5.5, GPT-6 Sol, GPT-6.1 Sol) list at or below the price of the model before them; the pairs are in Section 2. Second, 4 of the 22 rows are promotional prices with a stated end date, so the price in force today is not the price a budget should assume for next year. Three of those four are Gemini Flash models whose promotional price ends on 2027-01-01 (Section 2); Gemini 3.8 Flash, for example, is $0.75 / $3.75 from 2026-09-02 and $1.50 / $7.50 from 2027-01-01 (price-3a6d0688, price-1c477e6c).

2. Price moves inside the quarter

A model’s price is not fixed at launch. Each line below is two dated rows for the same model line, with the change computed from the two rows. The change column reads input / output.

The pattern in these rows is that vendors repriced an existing model (Terra, Luna, Sol) or launched a successor at the same or a lower rate. A scan of the price rows, applying the explorer’s rule that an active promotional row beats the standard row, finds 3 increases, all on 2027-01-01 when a promotional price ends: Gemini 3.6 Flash (price-1eae86b8 to price-123d4a64), Gemini 3.7 Flash (price-5de3a220 to price-5538411c) and Gemini 3.8 Flash (price-3a6d0688 to price-1c477e6c), each from $0.75 / $3.75 to $1.50 / $7.50. None took effect on or before 2026-09-30. Prices a vendor never published, or published without a start date, are not in that scan.

3. Independent agentic coding results: Terminal-Bench 4.0

Terminal-Bench is an agentic terminal and shell task suite. Version 4.0 is a different task set from 2.0 and 2.1 and is kept as a separate benchmark in the dataset. The rows below come from the public Terminal-Bench leaderboard (independent:src-tbench, retrieved 2026-09-11). The score is the share of tasks resolved, and the ± value is the interval the board published. The tokens column is the total tokens for the whole benchmark run, as the board recorded it.

The agent is part of the result. Claude Code (5 rows: Claude Fable 5.1, Claude Opus 5, GLM-5.3, Claude Opus 4.8, Claude Sonnet 5); Codex (3 rows: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna); Grok Build (2 rows: Grok 4.6, Grok 4.5); mini-SWE-agent (2 rows: Gemini 3.8 Flash, Gemini 3.7 Flash). A row therefore measures a model and its agent together, and nothing in this table isolates the model. Gemini 3.8 Flash at 19.1 and Claude Fable 5.1 at 57.9 were not run in the same agent.

Neighbouring rows are often not separable. 9 of the 11 adjacent pairs in the ordering above have overlapping published intervals, which means the board alone does not establish their order. The pairs whose intervals do not overlap are: Claude Opus 5 over GLM-5.3; GPT-5.6 Sol over Claude Opus 4.8.

Token use varies more than scores do. Claude Sonnet 5 used 21.6B tokens for 12.4%, while Claude Fable 5.1 used 2.7B tokens for 57.9% (obs-8d296c20, obs-f96b0b30). The only two dollar figures the dataset holds for these runs are $1.8k for the Gemini 3.8 Flash run (the source quote on independent:src-tbench) and $0.3k for the GPT-5.6 Luna run (the comparability note on obs-c812e3a0), so a full cost-per-resolved-task table cannot be built from this data. The token counts are a reminder that list price per million tokens is only half of a coding bill.

For orientation, here is each model’s list price in force on 2026-09-30 next to its independent score.

No ratio is computed from these columns, on purpose. The scores come from different agents, the runs used different token volumes, and a price per million tokens does not say how many tokens a task needed.

4. Do vendor-published figures match the independent board?

Vendors publish their own tables, often listing competitors. For every model with both a vendor-published and an independent Terminal-Bench 4.0 figure in the dataset, the table compares them. An asterisk marks a figure a vendor reported for a competitor’s model rather than its own.

4 of 7 models match to the decimal, and all 7 sit inside the independent board’s published interval. The largest difference is 2.1 points (Claude Fable 5.1 and GPT-5.6 Terra), and in each case it is smaller than the interval the board published for that model. The dataset records that the figures agree. It does not record why, and this page does not say.

5. The version trap: the same model on two Terminal-Bench versions

Vendor tables for earlier models cite Terminal-Bench 2.1, and newer tables cite 4.0. The dataset keeps them as separate benchmarks because they are different task sets. The table shows what happens if the numbers are read as one scale.

Every model in the table has a lower figure on 4.0, by up to 74.6 points. Claude Opus 5, for example, is listed at 89.1 on 2.1 (obs-cdd38868) and 51.8 on 4.0 (obs-fb96d50c). That is a change of ruler, not a change of model. A figure quoted without its version number is not comparable with anything.

6. Human preference for coding answers: LMArena

LMArena’s coding category ranks models by blind pairwise human voting. It is a different instrument from Terminal-Bench: voters judge answers, and no agent has to finish a task. The dataset holds 13 coding rows, all taken on 2026-09-03.

The spread from first to last is 33 Elo, and 4 of the 13 rows, the top row included, have an interval that overlaps the top row’s. The top of this list is a cluster, not a ladder. The two instruments also rank some models differently. Gemini 3.8 Flash is 4th of 13 on LMArena coding (1537, obs-98a1252c) but 8th of 12 on the independent Terminal-Bench 4.0 rows in this dataset (19.1, obs-16421340). Neither number is wrong. They measure different things.

7. SWE-bench in the dataset

SWE-bench Verified is a human-filtered subset of SWE-bench that tests resolving real GitHub issues end to end. The dataset holds 22 independent rows from the SWE-bench team (agent: mini-SWE-agent). Every one was evaluated between 2025-07-26 and 2026-02-26. None of the models released in Q3 appears in them, so SWE-bench Verified in this dataset describes earlier models. The eight highest rows:

Two other independent SWE-bench sources are in the dataset. Scale AI’s SEAL board lists SWE-bench Pro on a public set and on a private set. The dataset describes the private set as a held-out commercial set of real repositories that is not comparable with the public set, so the two columns below are side by side for reference, not a like-for-like gap.

All three models on both sets score lower on the private set (by 4.8 points for Claude Opus 4.6, 13.9 for Gemini 3.1 Pro and 22.5 for Claude Opus 4.5). The gap differs a lot by model, so a public-set score does not tell you where a model would land on the private set. Epoch AI’s 484-task subset of SWE-bench Verified, run with Epoch’s own harness, adds two more rows:

The two Epoch intervals overlap, so this subset does not separate Claude Opus 4.6 from DeepSeek V4 Pro.

8. DeepSWE: many vendor tables, no independent measurement

DeepSWE (version 1.1) appears in 33 observations, and 0 of them are independent. Every figure is a vendor’s own table. Vendors often print matching or near-matching numbers for the same model. For the 5 models listed by more than one vendor, the table shows who published which value. An asterisk marks a figure a vendor reported for a competitor.

The dataset does not record whether matching figures come from separate runs, so they are counted here as vendor claims and not as independent confirmation.

9. What the explorer says for a pure coding workload

The explorer’s coding preset asks whether any current model matches the default baseline, Claude Opus 4.6, for a workload that is all coding. Its answer for dataset version 2026-09-30.1, with comparison date 2026-09-30, is below. The explorer’s result carries this caveat from the dataset, which the compact embed below does not display: “These are API model comparisons. An agent product (Claude Code, Codex, Gemini CLI, Cursor…) adds tools, memory, orchestration and an execution environment; a model benchmark does not establish what an agent application can do.”

In the explorer’s words: “No model has enough comparable evidence to be called a baseline match for this workload. Closest candidates: Gemini 3 Flash (Preview) (55% covered), Kimi K3 (30% covered), Claude Fable 5.1 (30% covered).” For a coding-only workload the dataset therefore holds new scores but not enough comparable evidence to name a match for the baseline.

10. What this data supports, and what it does not

How to use this page

• Choosing a model for an API-driven coding agent: start from the LLM API pricing reference and the explorer, set your own mix of work, and treat any one benchmark as one input.
• Running Claude Code or Codex: the Terminal-Bench 4.0 rows in Section 3 are the only independent rows in the dataset that name those agents. Our Claude Code review, Codex CLI review and Codex token usage guide cover the products themselves.
• Watching cost per task: read the token column before the price column. Our GPT-5.6 Luna and Terra API review and DeepSeek V4 API review cover models that appear in the tables above.
• Comparing the products themselves: the AI coding agents hub lists the head-to-head comparisons.

Method and sources

All rows come from dataset version 2026-09-30.1 (published 2026-09-30, verified through 2026-09-30, flagged stale after 30 days). Prices are the rows in force on 2026-09-30, with promotional rows taking precedence. For a benchmark row the date is the evaluation date where the source gives one, and otherwise the date the source was retrieved, marked “retrieved”. The ± values are as published by each board. Overlap statements compare published intervals only and are not a significance test. Differences, percentage changes and counts are computed from the rows named beside them. The page was built on 2026-10-08 from dataset version 2026-09-30.1 and does not update when the dataset does, so check the explorer for current figures.