Changes since baseline #

Market changes #

The baseline is July 12, 2026; the new snapshot ends September 8, 2026, America/New_York. Later announced dates below are future terms known at the cutoff. This comparison does not isolate model improvements from changes in harnesses, evaluation methods, or this review’s source selection.

Release or change	What changed the decision	September action
GLM-5.3 and GLM-5.3-Flash	The text flagship requires reasoning; Coding Plan aliases for 5.2/5.1 route to 5.3. Flash is a separate multimodal model. Vendor and contributor terminal results justify fresh evaluation, with different experiment conditions.	Re-test the budget workflow; confirm endpoint/account compatibility and returned model. Compare retries and review effort. Apply the correct weight license when self-hosting.
GPT-6 Astra	A new escalation candidate adds async-tool and steering support. Standard API rates are $10 input/$50 output per MTok; long-input and Fast premiums can change the bill.	Shortlist Astra with a compatible local Codex route for difficult tasks. Verify client implementation, permissions, and account access.
Claude Fable 5.1	Released September 1; standard $10/$50 API rates persist, while cache reads fall from $1 to $0.25 per MTok. Thinking/tool-choice behavior and retention/fallback conditions differ.	Re-evaluate July’s Fable choice on hard tasks and repeated context. Review middleware and the exact privacy route before switching.

Sources: GLM migration, plan routing, Flash card, Astra features, Astra rates, Fable specifications, migration, retention and fallback. Claims CR01–CR05, CGL01–CGL08, CGL20 and CO04–CO06 contain scoped support.

Opus 5 adds a lower-priced Anthropic starting candidate. Gemini 3.8 Flash adds a greenfield/agentic pilot candidate, with introductory Developer API prices; Antigravity managed-agent billing is unverified here. DeepSeek V4-Pro is now GA and has peak/off-peak rates and a documented Codex integration, with Responses semantic differences. Sol’s current $4/$20 promotional API rates replace July’s recorded $5/$30 snapshot. Opus release, Gemini release, DeepSeek updates, Sol rates.

Terminal-Bench 4.0 changes tasks and resources and requires reruns. September’s Astra/Codex and Fable/Claude Code observations are too close and incompletely specified to select a winner; they are not comparable with earlier versions as a progress series. Maintainer notice, contributor results.

Copilot’s temporary Business/Enterprise credit uplift ended September 1; new PR-approval preview changes authority, not demonstrated review reliability. Cursor’s third-party token fee and Fable retention exception matter to route selection. Copilot billing, approval preview, Cursor billing, model governance.

Rechecked recommendations #

Keep an established Sol/Codex workflow for everyday work when its accepted results and costs remain satisfactory. Opus 5 is an alternative to test, not a proved universal replacement. The new frontier models justify escalation trials rather than a mandatory migration. Local 8/16/24 GB routes remain pilots: theoretical weight storage does not establish context capacity, latency, tool reliability, or production suitability.

Human ownership, tests, review, and bounded authority remain the operating recommendation. Existing privacy-constrained teams still need a product/model/route-specific review; a vendor’s no-training default does not establish zero retention. These judgments are conditional and are linked to disconfirming criteria in the recommendations.

Corrections #

These are corrections to interpretation or presentation, separately identified from new market events. July’s files are preserved.

• Price arithmetic: July’s “exactly half” Sol/Fable summary did not follow from its own $5/$30 versus $10/$50 table: output was 60%, not 50%. September records current rates and labels workload calculations as estimates. The illustrative Fable cache scenario is 20% cheaper under stated assumptions; it is not a task-cost result.
• Privacy scope: July’s blanket no-training summary conflicted with its DeepSeek row. September uses exact routes and leaves unsupported API-wide conclusions unknown. A personal-data policy is not automatically an API contract.
• Productivity condensation: July’s full report already included METR’s February follow-up. September restores its uncertainty and selection caveats to the condensed interpretation; it does not claim that follow-up is new research discovered after July.
• Comparability: Historical mixed-harness percentages are no longer used to imply a common ranking. New observations retain dataset version, evaluator, uncertainty, and missing settings.
• Attribution: Researcher stamps identify who produced an edition; they cannot isolate researcher-model effects in live web reviews.

The approved methodology review documents these internal baseline findings. Current external terms were checked afresh; they are not inferred from July.

Methodology changes #

Methodology 1.1.0 is a minor revision: it preserves the research question, comparison framework, and 13 sections. It adds conditional setup recommendations, linked source/claim records, explicit sampling and comparison rules, exact data schemas, a change log, and editorial review. Historical v1.0.0 data retain their own adapter and are not retroactively certified against v1.1.0.

Final policy review also separates Z.ai API commitments from the unresolved individual Coding Plan mapping. The broad GLM pilot recommendation therefore has low confidence and requires an approved route. This is an evidence-scope finding, not a claim that the plan trains on code. API terms, DPA, subscription terms.

The edition contains a readable register, two evidence ledgers, seven canonical datasets, and generated recommendation tables shared across reading layers. Structural validation checks consistency; author and parallel peer checks assess support. No original comparative experiment or independent human review was performed.

Open questions #

The next useful evidence is accepted-task performance on representative repositories, including failed attempts, review/rework, total bill, and time. Specific watchlist items are:

• Astra versus Fable 5.1 with complete harness, budget, attempts, and fallback records.
• GLM-5.3 account/protocol interoperability and actual task economics; Flash after its announced promotion ends September 9 at 24:00 UTC+8.
• Fable enterprise retention eligibility and migration behavior on existing sessions.
• Qwen3.8-27B at an actual supported quantization on 24 GB hardware; Kimi K3, MiniMax M3, and Flash-Next where serving/license constraints fit.
• Gemini introductory-price expiry and Antigravity’s separate managed-agent bill; Sol after its promotional guarantee.
• Independent review-agent defect detection and reliable bounded background work, including failures and human attention.

These are evaluation questions, not scheduled monitoring or endorsements of untested setups. See research limits and canonical records.