State of Agentic Coding — September 2026 #

Snapshot: September 8, 2026. Research: September 9 UTC, beginning September 8 in New York. Researcher: GPT-6 family (gpt-6; exact backend variant unknown), Codex desktop with three parallel research workers. Methodology and data schema v1.1.0. Baseline: July 12 edition.

This edition is an evidence review, not an original model benchmark. Its recommendations are starting points for specific operating conditions. The important September decision is to re-evaluate the complete workflow: the three requested releases—GLM-5.3, GPT-6 Astra, and Claude Fable 5.1—change the available capabilities, bills, and controls. They do not establish one winner for every repository. Read the decision tables, changes since July, and source register alongside this report.

1. Landscape overview #

Fable 5.1 replaces Fable 5 as Anthropic’s current escalation candidate for demanding work. Opus 5 arrived July 24 and remains Anthropic’s recommended starting point for most workloads. Fable’s stronger positioning comes with model-specific migration and retention requirements. This supports trying 5.1 where July recommended Fable, while measuring whether its extra capability justifies its bill. Fable model guidance, Opus 5 release.

Astra is an additional frontier candidate, with documented support for asynchronous tools, steering during a task, and reasoning-effort changes that preserve cache. These are useful integration options for long jobs; the application still has to manage pending tools and interruptions. A model upgrade alone does not deliver those workflow benefits. Astra guidance.

The open roster changes materially: GLM-5.3 is the updated text flagship, while GLM-5.3-Flash is a separate multimodal model with a new base and different economics and license. Z.ai’s release notes place these entries on August 18 and August 26; these are documentation dates, not verified first announcement times. Z.ai release notes

The useful GLM change is more than a name swap. Z.ai reports a substantial Terminal-Bench 3.0 improvement from its 5.2 baseline, with detailed harness settings, but this remains a vendor experiment rather than evidence that it replaces Astra or Fable for every task. Existing integrations also change behavior: reasoning cannot be disabled, and Coding Plan requests naming 5.2 or 5.1 now route to 5.3. Log the returned model, explicit effort, endpoint and client before interpreting a before/after result. Model card, migration documentation, plan routing

Qwen3.8-27B, Kimi K3 and MiniMax M3 broaden the candidate roster. Their serving and license conditions are recorded separately; Flash-Next is an experimental architecture preview. MiniMax M3 is added coverage, not presented as a September launch. Qwen model, Kimi model, MiniMax model.

Google’s September 2 Gemini 3.8 Flash release adds a lower-priced candidate for agentic work, including the default Antigravity managed agent. Flash Cyber has restricted defender access and should not be treated as the generally available model. DeepSeek’s hosted V4-Pro reached GA August 13; its July 31 Flash update and new Responses integration also make July’s API snapshot stale. Gemini release, DeepSeek change log.

Discovery included the larger providers and open ecosystem before revisiting July’s winners. This report gives most attention to candidates that alter a decision; it does not imply excluded products disappeared. The date cutoff uses America/New_York; these UTC September 9 checks occurred on September 8 locally. Undated documentation records that observed state, not an hour-exact historical snapshot. Announced future price changes are described as announced terms, not as events that already happened.

2. Frontier model comparison #

The new capability evidence describes complete systems. Snorkel’s contributor presentation of Terminal-Bench 4.0 lists Astra with Codex at 58.2% (reported margin ±2.8), Fable 5.1 with Claude Code at 57.9% (±3.8), and GLM-5.3 with Claude Code at 41.8% (±3.2), all at max effort. The small Astra–Fable gap is insufficient evidence of superiority. The source does not define these margins as 95% confidence intervals, and missing run metadata keep the observations outside ranked comparison groups. Contributor leaderboard.

Terminal-Bench 4.0 changes the experiment: 66 tasks remain, the timeout is 8 hours, resources are recalibrated and 8 tasks were removed. The maintainer reports 19 task fixes, while Snorkel describes 20 revisions. The maintainer requires rerunning trials. Earlier versions therefore cannot form one continuous score series with September results. Maintainer explanation, versioned release.

For difficult debugging and refactoring, put Astra/Codex and Fable 5.1/Claude Code on the evaluation shortlist. Prefer the established client if changing it would add migration cost, then challenge that preference on actual failing tests, architecture constraints, and reviewer effort. GLM-5.3 is a budget alternative where its lower bill survives retries and review. A benchmark shortlist is not evidence that any of these is best for a specific Java service, TypeScript UI, Python pipeline, or unfamiliar C++ subsystem.

Anthropic’s launch evidence is encouraging but must retain its origin. Its reported Fable 5.1 results include production safeguards and fallback behavior; selected early-access stories are not independent replication. Its own Terminal-Bench-Science note explains that reruns differ within reported uncertainty, and its OSWorld results use an August task revision. Those qualifications rule out subtracting a July score from a September score and calling the difference model progress. Fable launch evaluation, evaluation footnotes.

The practical comparison unit is a recorded model, client version, effort setting, tools, permissions, and task. Keep accepted outcomes, rejected attempts, elapsed time, token charges, and human review separate. Missing latency or reliability measurements remain unknown in the datasets. No synthetic overall score combines unrelated benchmark percentages, policy labels, and prices.

3. The harness ecosystem #

August VS Code releases added Agent Plugins 1.0 and continuation of external Claude/Copilot sessions. Zed exposes an ACP route with provider authentication and billing owned by the agent. This documents integration portability; complete session interchange and equal tool behavior remain unmeasured. VS Code, Zed.

OpenCode documents hosted and local providers, with distinct adapters for Responses and Chat Completions. Cursor supports native Fable use subject to model retention opt-in; Claude Code supports local and web execution with different data paths. This is useful workflow coverage, without an independent comparison of context retrieval, recovery or browser reliability. Aider, Cline and Devin were considered during discovery, but the reviewed material does not support a fresh across-client ranking. OpenCode providers, Cursor model governance, Claude Code data paths.

Long-session capability needs observable state: an updated plan, durable decisions, a record of attempted fixes, the commands actually run, and a useful failure report. Context windows describe capacity; they do not measure retrieval quality or whether the agent remembers the right invariant. Compaction and memory should preserve decisions and evidence, not just conversation volume. These are evaluation criteria proposed by this review, rather than features claimed for every client.

Native interfaces can implement new model behavior sooner than compatible endpoints. Conversely, an open client can reduce migration effort if the provider changes. Neither property guarantees a better patch. Test recovery from a failed command, exhausted context, interrupted run, and tool timeout before trusting a workflow with a long unattended task.

4. Project size and maturity #

For greenfield and UI work, use a candidate that can inspect the rendered result and run the application’s real checks. Ask for a demonstrable user flow, not only a screenshot or generated component count. Gemini 3.8 and Anthropic’s newer models have vendor demonstrations of visual and application work, which justify evaluation, not a measured greenfield ranking. Gemini examples, Opus examples.

In a growing project, begin with a bounded implementation task, an explicit acceptance condition, and an existing test command. Use Sol or Opus 5 as a lower-priced starting candidate, and escalate after diagnosing why it failed. Repeated failures can be caused by missing context, an unavailable service, ambiguous requirements, or an incorrect test—not only insufficient model capability. This proposed escalation rule should be checked against the team’s own results.

For a large or legacy repository, map the affected architecture before changing it: ownership, data contracts, migrations, deployment assumptions, and cross-service callers. A 20,000-line repository with weak tests can be harder to change safely than a much larger familiar codebase with narrow interfaces. No line-count threshold in this edition predicts reliable autonomy. For maintenance and security fixes, preserve the failing case, verify the repair, and inspect adjacent behavior. Familiarity and reviewer comprehension belong in the evaluation alongside test passage.

For long sessions, require checkpoints at meaningful boundaries and compare total work accepted after review. An overnight agent that produces an unreviewable diff has not demonstrated a productivity gain. A shorter, verified change can be more useful even if another setup has a higher terminal benchmark score.

5. On-prem and self-hosted #

For a local pilot, use memory capacity to narrow candidates, then measure the agent. The 8/16/24 GB rows propose Qwen3.5-4B, Qwen3.5-9B and the new dense Qwen3.8-27B at explicitly hypothetical precisions. Simple weights-only arithmetic leaves room for investigation; it establishes neither a working quant nor usable context. Quantization scales, mixed-precision components, caches, recurrent state, activations and concurrent sessions add memory. A 64 GB unified-memory machine shares its pool with the OS and applications; it is not a 64 GB dedicated GPU. The edition has no measured latency or accepted-task success rate for these exact deployments. Start with scoped changes and fixed tests; measure peak memory, valid tool calls, accepted output, retries and review effort before buying hardware or permitting unattended work. Qwen 4B, Qwen 9B, Qwen3.8-27B

The arithmetic in claim CGL18 uses total parameters, not activated MoE parameters: 28B weights at four bits require about 14 decimal GB before overhead; GLM-5.3 at a hypothetical four bits needs about 376.5 GB for weights alone. GLM-5.3-Flash is also a server-scale model despite its name. No table row labels theoretical storage as measured VRAM. GLM card, Flash card.

Offline operation is an end-to-end property. The model, serving process, client, embeddings, telemetry, dependency fetches, and external tools all need to respect the boundary. A local inference endpoint can still sit behind a client that sends repository content elsewhere. Memory fit alone establishes neither acceptable latency nor reliable tool use. Record time to first token, generation speed, successful tool calls, usable context under load, and accepted tasks on the actual hardware before moving a team to that route.

No hardware purchase price or electricity tariff was assumed. A useful local-cost model includes acquisition or rental, depreciation period, power, utilization, maintenance, and operator time. Without those inputs, comparing zero API fees against a hosted token bill would conceal the local system’s costs.

6. Pricing #

The canonical pricing dataset records direct API and subscription routes separately. Astra lists $10 input and $50 output per million tokens. Sol currently lists promotional $4/$20, available at least through November 21. Astra requests above 272K input tokens receive higher rates for the whole request; Fast also has a premium. Those conditions matter for long-context agents and prevent a single output-price bar from describing the bill. Astra rates, Sol rates.

Fable 5.1 keeps the $10/$50 base rates and cuts cache reads from $1 to $0.25 per million tokens. Five-minute writes cost $12.50; one-hour writes $20. The practical gain depends on cache reuse and output volume. Opus 5 lists $5/$25 and Sonnet 5 $2/$10; the latter is a June release, not a new September contender. Cache rate table, Sonnet specifications.

Illustration, not measured task cost: aggregate 1M uncached input tokens, 1M five-minute cache-write tokens, 10M cache-read tokens, and 0.1M output tokens. Assume every read hits, every request stays below the long-input threshold, and exclude tools, tax, and labor. Applying input × rate + writes × rate + reads × rate + output × rate gives $30 for Fable 5.1 versus $37.50 for Fable 5, a 20% reduction. The same token counts give Astra $37.50 and Sol $15. Different tokenizers, outputs, retries, and acceptance rates prevent interpreting that arithmetic as a quality-adjusted comparison. Claim CR15 links the inputs and formula.

Gemini 3.8 Flash’s $0.75/$3.75 introductory rates end December 31 under the current announcement; January’s announced base rates double. Cache storage and grounding are additional charges. Google also warns that harder work may consume more tokens. DeepSeek announced peak/off-peak pricing effective August 16, and a direct public-page fetch resolved an initial extraction failure. V4-Pro peak rates are $1.32 input/$3.96 output per MTok; off-peak halves them. Peak hours are Monday–Friday 01:00–04:00 and 06:00–10:00 UTC. The full matrix also records Flash rates. Gemini rate schedule, DeepSeek update, current pricing.

Recalculate allowances before upgrading models: Copilot’s temporary Business/Enterprise credit uplift ended September 1. Cursor Teams adds a per-token fee to third-party prices, including cache and BYOK. Copilot, Cursor.

Z.ai lists GLM-5.3 at $1.40 input/$4.40 output per MTok, with cache reads $0.26. Flash list rates are $0.15/$0.50, with a separate expiring discount. Coding Plan credits are token-weighted and constrained by both five-hour and weekly caps; older fixed-task interpretations should be retired. The starting $18/month advertisement does not establish Pro/Max checkout prices or an unrestricted API allowance. Z.ai API rates, Coding Plan.

For actual procurement, calculate total model and tool charges across successful and failed attempts divided by accepted tasks, then report review minutes alongside that number. Only monetize those minutes when a labor rate is supplied. Subscription quotas, shared credits, and API balances are different purchasing units; a plan price does not establish a stable number of coding tasks.

7. Client and IDE compatibility #

The compatibility matrix distinguishes native, BYO-key, and compatible-endpoint support, along with execution surface and account route. All entries are documentation-based unless explicitly stated otherwise; this review did not exercise paid accounts or run a client interoperability suite.

Fable 5.1 has concrete breaking API changes. Forced tool_choice values any and tool return errors. Earlier Claude models cannot consume its thinking blocks, so a fallback may lose reasoning context. Editing the prefix can invalidate later blocks, with enforcement dependent on account vintage and settings; official Claude clients handle prefix preservation. Check custom routers, prompts, and middleware before a model-name substitution. Migration changes.

Astra’s async calls and steering likewise require matching client behavior. DeepSeek documents a Responses route for Codex and explains that switching authentication routes changes which session-history group is visible. DeepSeek also documents ignored fields and built-in tools, including background/state parameters and built-in MCP; its developer role becomes user. Inspect semantic differences, not just whether the request succeeds. Compatibility details. Astra API guidance, DeepSeek integration.

GLM-5.3 now requires reasoning; disabling it fails. Z.ai lists several protocols but adds an account-history restriction for previous Coding Plan subscribers, while its plan page names Claude Code, Cline and OpenCode. That route ambiguity belongs in the matrix: confirm the intended endpoint and account before switching. Zed ACP connects external agents, but their own billing and configuration still apply. GLM migration, Coding Plan routes, Zed external agents.

8. Privacy and training #

Fable 5.1 changes the privacy decision as well as model selection. Copilot and Cursor explicitly exclude it from ordinary provider-ZDR defaults. Check enterprise eligibility or retention opt-in on the actual route. Copilot hosting, Cursor governance.

OpenAI API no-training defaults coexist with abuse logs and endpoint-specific state retention. Codex Business no-training is documented, while exact Business retention and consumer Codex settings remain gaps in this review. Claude Code separates consumer data-sharing choice from commercial defaults; local transcripts, feedback, web execution and model-specific exceptions add layers. Cursor BYO keys still pass through Cursor. OpenAI API data, Codex plan terms, Claude Code data, Cursor data use.

DeepSeek requires a scope check: its personal-data policy permits training with opt-out but excludes downstream end-user processing rules; platform terms repeat the boundary. This review does not establish an API-wide no-training promise. A stateless endpoint is not such a promise. Personal-data policy, platform terms.

Google’s Gemini Developer API illustrates why tier names need precise interpretation. Paid-service content is excluded from product improvement, but abuse logging still exists. Unpaid-service treatment differs by region; EEA, Switzerland, and UK users receive the paid-service data terms even on unpaid quota. Search grounding has its own 30-day retention. Record the actual billing-linked project and tools instead of borrowing the Gemini consumer app’s policy. Gemini terms.

This edition does not certify regulatory compliance or infer contractual protection for every enterprise/API plan. The matrix records scoped commitments and unknowns. The practical review is whether the chosen model and intermediary can receive this repository’s data under the organization’s requirements, including retention, human access, deletion, residency, and external tools.

Z.ai’s enterprise/developer API terms restrict content use for improvement without explicit agreement, and its API DPA promises no server storage of submitted/generated content. These commitments have defined contractual scope. Individual Coding Plan subscriptions are separately restricted to supported tools; the reviewed terms do not explicitly settle whether that route receives the API protections. Keep its training and retention fields unknown, and require an approved data path before use with restricted repositories. API terms, API DPA, Coding Plan subscription terms.

9. Security and authority #

Astra monitoring in Codex can pause tasks asynchronously, after a triggering action. Sandbox, network policy and review remain necessary. Copilot PR-approval preview separately changes review authority: opt-in approval can satisfy merge requirements. Neither announcement measures independent review recall or reliable unattended operation. Codex, Copilot.

Endor’s September 4 Fable 5.1 with Claude Code result separates functional success from joint functional and security-test success: 87.2% FuncPass versus 37.4% SecPass. Endor is a security vendor; its method adjusts for cheating and infeasible cases, and exact September run metadata are incomplete. GitInject’s real CI/CD experiments separately show why credentials and trusted repository instructions need explicit boundaries. Those older configurations do not prove every current model has the same vulnerabilities. Security leaderboard, method, GitInject.

Separate permission to inspect, edit, execute, access the network, read credentials, approve a pull request, merge, and deploy. Repository text, retrieved web pages, tool results, plugins, and dependencies can introduce untrusted instructions. A capable model still needs enforceable boundaries and a reviewable record. Use isolated execution where appropriate, narrowly scoped credentials, explicit network controls, rollback, and human ownership of consequential changes. These are proposed operating controls, not a claim that any client eliminates prompt injection.

Keep the reviewer able to challenge the implementation and its tests. Changing the reviewer model may provide another perspective, but does not establish statistical independence: both agents may inherit the same assumptions, context omissions, or benchmark incentives. Test passage should be supplemented by security and architecture checks suited to the change.

10. Measured vs perceived productivity #

There is no supported universal productivity multiplier. METR’s early-2025 slowdown remains an observation for its original setting, while its February 2026 follow-up showed possible speedups with intervals crossing zero and serious selection and timing problems. Its May survey found larger perceived gains but used a convenience sample with roughly 2% email response. A positive randomized field result from 4,867 developers concerned an older completion assistant. None identifies the effect of the three new models. METR follow-up, survey, field experiments.

A July 29 preprint with 54 students found better initial task performance but weaker understanding and little transfer to extension work. It used GPT-4.1 with Aider, and the extension phase still permitted a chatbot. This supports checking whether an owner can explain and maintain the code, rather than a claim about September models’ effect on professional productivity. Primary study.

The first-hand sample is small and mixed. A student describes advantages and concrete failures in both Astra and Fable 5.1, followed by gains from human feedback; its exact post date is unresolved. A separate microtest’s only Fable failure disappeared under lenient output grading. Tim Sehn’s August database work favored Sol/Codex for implementation and Claude for ambiguity. David Vujic’s June account says improved review flow avoided a bottleneck. These four authors across Reddit, a company blog and personal blogs illustrate workload variation, not consensus. ML account, microtest, database account, review-flow account.

The Reddit account has inconsistent relative dating in the extracted page, so its exact publication day is uncertain; crossposts count once. No usable GLM-5.3 first-hand pattern was confirmed. No recurring pattern is claimed from these heterogeneous accounts, even though the sample spans several venues. The complete selection and exclusion record is in research-log.md.

For a team trial, predefine useful outcomes and include abandoned attempts. Track accepted task completion, review/rework, incidents or regressions, and elapsed time under the same task mix. PR count, lines generated, runtime, and self-reported speed describe different phenomena. The report has no new causal estimate for Astra, Fable 5.1, or GLM-5.3 on production engineering teams.

11. Vendor risk and portability #

Open weights now require model-specific license review. GLM-5.3 has a custom security-review condition for large Model as a Service operators; GLM-5.3-Flash retains MIT. Qwen3.8-27B is designated Apache 2.0, whereas Flash-Next’s custom terms specifically address commercial AI Work Assistants and MaaS. Kimi K3 and MiniMax M3 also carry distinct revenue, notice, attribution or agreement conditions. These are procurement differences, not benchmark results; do not extend an older model’s permissive license to its successor. GLM-5.3, Flash, Qwen Flash-Next, Kimi, MiniMax

Coding Plan aliases make version labels especially important: a request still naming GLM-5.2 or 5.1 can reach 5.3. Preserve returned model identifiers where exposed and avoid attributing a changed result to the same named system. Plan routing.

Fable’s default retention and eligibility rules, Astra’s long-input pricing, and Gemini/Sol promotional terms can change the preferred route without any change in coding quality. Distinguish a preview or phased entitlement from account access actually confirmed. Do not assume a model available in one vendor’s app is available in the same region, client, or enterprise contract elsewhere. Fable access, Astra rates, Gemini rates, Sol promotion.

Portability requires more than an endpoint setting: prompts, thinking-state formats, tool schemas, skills, authentication, session history, billing, and approved data routes can differ. Preserve a minimal reproducible task and its acceptance checks for migrations. IP and commercial-use terms should be checked on the exact weight license or service contract; this edition does not infer an indemnity from a paid subscription or give a general regulatory forecast.

12. Team workflows #

Parallel work is useful when tasks have clear ownership and can be integrated without competing edits. Assign a boundary, acceptance checks, and a handoff record; keep one owner accountable for integration and review. Limit work in progress when the review queue grows. Background execution should stop at a defined artifact or decision, with failures and remaining uncertainty visible.

Start unattended trials on bounded, reversible tasks with clear checks. Increase scope only after observing outcomes, including failed runs. A vendor’s multi-hour demonstration can establish that a workflow is possible; it cannot establish a dependable unattended production success rate. This edition recommends a staged workflow trial, not automatic delegation of merge or deployment authority. Fable launch evidence and selected customer reports.

For teams, keep the model evaluation and process evaluation connected. If more generated patches increase integration delay, investigate task overlap, reviewer capacity, and acceptance criteria before buying more parallel capacity. If a stronger model reduces review effort on costly tasks, a higher token rate may be justified. Neither relationship was directly measured by this research run.

13. Recommendations #

These choices are conditional evaluation defaults. Medium confidence means the documented terms and relevant evidence justify a starting point, while repo-specific quality or economics remain unresolved. Low confidence marks missing comparative evidence or an untested local setup. No recommendation treats a vendor announcement as independent proof of superiority.

Recheck the established workflow before migrating: a reliable Sol/Opus setup can remain the everyday choice; use Astra and Fable 5.1 for difficult work where an observed improvement pays for the extra cost. GLM-5.3 deserves a fresh budget evaluation, including endpoint and hosted-data constraints; weight licenses matter separately for self-hosting. Keep newer or thinly measured alternatives in the watchlist until their exact setup passes the team’s acceptance checks.

The watchlist includes GLM Flash after its promotion, Qwen3.8-27B on measured local hardware, Kimi K3/MiniMax M3/Flash-Next under the relevant serving and license constraints, Fable enterprise retention eligibility, and independent review-agent outcomes. Recheck Sol and Gemini promotional terms before committing a longer budget. The change log records the specific evidence that would justify revisiting these choices.