Claude Code weekly limits drop 17% on 14 September — Keep near-sonnet performance with Qwen3.8-27B & GLM-5.2 on Regolo

On 13 September 2026 at 23:59 PT the temporary 50 % increase to Claude Code weekly limits expires. From 14 September the permanent allowance sits 25 % above the old baseline — roughly 17 % less capacity than most Max and Pro users have been consuming for the last four months.

This guide shows a concrete alternative that European teams are already running: pair the strong open models available on Regolo (especially Qwen3.8-27B and GLM-5.2) with a proper agent harness and Brick as the semantic router. The combination delivers performance close to Sonnet/Opus on the majority of real coding-agent workloads, with predictable cost, Zero Data Retention, EU infrastructure, and no weekly hard stop — including weekends.

---

---

The models that actually close the gap

Two models currently available on Regolo stand out for agentic coding:

Qwen3.8-27B
Dense 27 B model. On Artificial Analysis it scores in the same band as much larger systems on many coding and agentic evaluations (Intelligence Index figures in the mid-to-high 30s to low 50s depending on reasoning setting). It is particularly strong for its size on SWE-style and terminal-agent tasks. Pricing on Regolo (pay-as-you-go): €0.50 / €2.10 per 1 M input/output tokens. Included in all subscription plans.

GLM-5.2
Leading open-weights model on several Artificial Analysis leaderboards at launch (Intelligence Index in the high 30s to low 50s depending on variant and date). Excellent long-context and agentic behaviour. Pricing on Regolo: €2.00 / €5.20 per 1 M. Also included in subscriptions.

Neither model is a perfect drop-in replacement for the absolute frontier closed models on every single hard reasoning task. With a good harness, however, the majority of coding-agent turns (file exploration, focused edits, test writing, simple refactors) land in the same quality band while costing dramatically less.

The harness matters more than the model name

Raw model calls are not enough. The difference between “good open model” and “near-Sonnet agentic performance” is the surrounding system: tool use, memory, verification loops, and routing.

Brick as pure semantic routing

Brick is used only as the routing layer, It classifies each request by capability and complexity, then forwards it to the cheapest model that can still handle the turn well. (arxiv paper)

Typical pattern below:

• Easy exploration / simple edits → Qwen3.8-27B or smaller
• Standard implementation → Qwen3.8-27B
• Deep planning or large refactors → GLM-5.2

Brick itself (brick-v1-beta or brick-complexity-pro) sits in front and adds almost no latency, this is the same cost-control pattern used in earlier Regolo multi-agent tutorials.

---

Cost reality under the Boost subscription

Getting started with Regolo

1. Create a Regolo API key.
2. Point your agent harness at https://api.regolo.ai/v1.
3. Use the Hands-On AI Engineering patterns (or your existing OpenCode / Cline / custom ReAct loop).
4. Put Brick in front of a small pool: Qwen3.8-27B + GLM-5.2 (plus a cheap explorer if desired).
5. Run the same multi-file refactor + test tasks you normally run on Claude Code and measure quality, tokens, and wall-clock time.

Because everything is OpenAI-compatible, the migration is usually a base-URL and model-list change, not a rewrite.

---

---

Weekend and residency side-effects

There is no separate capacity pool that shrinks outside US business hours. The same models and the same routing rules are available Saturday and Sunday. All inference stays in European data centres with Zero Data Retention by default — useful for any team that has GDPR, customer-contract, or internal residency constraints.

Actionable next steps

1. Open the Regolo playground or create an API key.
2. Point a copy of the Hands-On AI Engineering agent examples at Regolo.
3. Add Brick as the semantic router in front of Qwen3.8-27B and GLM-5.2.
4. Re-run three or four of your real Claude Code workloads and compare.
5. Decide whether the remaining hard cases still need a closed model or whether the open stack already covers 80–90 % of the volume at 1/10th the cost.

The weekly limit change on 14 September is real. The productivity drop does not have to be.

With Qwen3.8-27B, GLM-5.2, a solid agent harness, and Brick doing pure semantic routing on Regolo, most teams can keep working at full speed — with transparent European pricing, no weekly ceiling, and data that never leaves the EU.

---

FAQ

When does the Claude Code 50% weekly boost end?
The temporary 50% increase to Claude Code weekly limits ends on 13 September 2026 at 23:59 PT. From 14 September the permanent allowance is 25% above the pre-May baseline, which is approximately 17% lower than the boosted level most users have been running on.

Do the 5-hour session limits also change?
No. Only the weekly usage pool is reduced. The rolling 5-hour session limits remain the same.

Which Regolo models are recommended as alternatives for coding agents?
The two strongest options currently available on Regolo for agentic coding are Qwen3.8-27B and GLM-5.2. Qwen3.8-27B is a dense 27B model that scores competitively on Artificial Analysis coding and agentic evaluations for its size. GLM-5.2 is a top open-weights model with strong long-context and agentic performance.

How do Qwen3.8-27B and GLM-5.2 compare to Claude Sonnet or Opus?
They are not identical to the absolute frontier closed models on every hard reasoning task but with a proper agent harness and routing, however, they deliver performance close to Sonnet/Opus on the majority of real coding-agent workloads (file exploration, focused edits, test writing, standard refactors).

What is Brick used for in this setup?
Brick is used purely as a semantic routing model. It classifies each request by capability and complexity, then routes it to the most cost-effective model in the pool (typically Qwen3.8-27B for standard turns and GLM-5.2 for harder planning or large refactors).

How much does this cost on the Regolo Boost plan?
The Boost plan is €89 per month and includes up to 50 million tokens per day (approximately 1.5 billion tokens per month). Qwen3.8-27B and GLM-5.2 are included in the subscription. At full utilisation the effective blended cost is roughly €0.059 per million tokens. Even at more moderate volumes the cost remains dramatically lower than typical Claude Sonnet or Opus list prices (often 1/10 or better on pure token spend).

Are the models available on weekends?
Yes. There is no separate capacity pool that shrinks outside US business hours. The same models and routing rules are available every day of the week.

Does Regolo keep any of my data?
No. Regolo operates with Zero Data Retention by default. All inference runs in European data centres and complies with GDPR requirements.

How do I switch my existing agent to Regolo?
Most OpenAI-compatible harnesses (OpenCode, Cline, custom ReAct/LangGraph agents, or the Hands-On AI Engineering examples) only require changing the base URL to https://api.regolo.ai/v1 and supplying a Regolo API key. Then configure Brick in front of a pool containing Qwen3.8-27B and GLM-5.2.

Is this a complete replacement for Claude Code?
For many teams it covers 80–90% of daily agentic coding volume at far lower and more predictable cost. A minority of extremely hard reasoning tasks may still benefit from a closed frontier model; the rest can run on the open stack without hitting weekly limits.

---