In partnership with
ShanghAI — Same Model. 10x the Invoice.
Hey Lobsters, welcome to this new edition of ShanghAI.
Last month we ran MiniMax's numbers: 76.7% task success at $1.82 per completed task. Claude Code: 63.3% at $18.34. Same 30-task suite. Same Kimi K3 model underneath both harnesses.
Six weeks later, the board moved. Claude Code's October wave (v2.1.280 to v2.1.296) made Opus 5.5 the default on every paid plan, at a $4 / $20 per million token list price.
MiniMax shipped 13 CLI releases in 15 days, launched M Plan, and moved its default to M3.1 Flash at $0.15 / $0.60.
Moonshot shipped Kimi K3.1 Pro, which scores 65.7% on Terminal-Bench 4.0 — within one point of Opus 5.5 — for roughly 45% of the token price.
The invoice did not change. It got clearer.
Here is the uncomfortable part: $18.34 is not Anthropic's model tax.
FrontierHarness holds the model fixed — Kimi K3 — and varies only the harness.
Claude Code's harness runs up roughly 10x the token bill to finish the same work. That is the line item your CFO should see.
📢 OUR PARTNER
ChatFeatured new GEO tool: Statistics
From invisible to first in ChatGPT.
Bob's Your Uncle has spent twenty years helping challenger brands win their categories. But when buyers asked ChatGPT about their own category, the agency didn't show up.
With ChatFeatured, their AI visibility went from 7% to 22% in 90 days, and they went from invisible to first in ChatGPT.
They're not the only ones:
Vessel: AI visibility up 2,200%+
SCOJO: AI visibility up 160%
WikiSleep: brand mentions up 113%
ChatFeatured finds the questions where AI leaves you out, writes the content that gets you cited, and reports the lift every week. Your team decides what ships.
Most customers see their first citations inside 7 days.
⚙️ HOW IT WORKS
THE 3 PRINCIPLES BEHIND THE INVOICE
Forget the model leaderboard for a minute. Three mechanics decide what you actually pay per merged pull request.
The harness is the invoice. Terminal agents re-read the world on every turn: git diffs, directory trees, test output, failed stack traces. Claude Code favors completeness. MiniMax favors compaction. Version 0.6.12 now keeps your latest request plus failed paths during compaction, so the agent stops re-deriving work it already did. That is a direct token cut, not a prompt trick.
Token price is a second-order variable. Anthropic says Opus 5.5 costs 40% less to run than Opus 5, and its list price is $4 / $20. That is still about 27x MiniMax M3.1 Flash on input. But if the harness runs up 10x the token bill, the cheap model loses its advantage. Price per task, not per token.
Benchmarks are marketing until reproduced. The official FrontierHarness board lists 12 harness configurations. MiniMax Code is not one of them. The 76.7% figure is a vendor-run result. It may be true. It is not independent. Demand a rerun on your repos before you sign anything.
📊 THE MATRIX
WHO IS WINNING AT THE END OF OCTOBER 2026?
Criteria | Claude Code | MiniMax Code 0.6.12 |
|---|---|---|
Default model | Opus 5.5 since v2.1.280 | M3.1 Flash via M Plan |
API price / 1M tokens | $4 input / $20 output | $0.15 input / $0.60 output |
Cost per completed task | $18.34 (FrontierHarness) | $1.82 (vendor self-report) |
Pass rate, 30 tasks | 63.3% (independent) | 76.7% (vendor-run) |
Median time per task | 9m 38s | 4m 33s |
Entry plan | Pro $20/mo ($17 annual) | M Plan Go $220/yr (~$18.33/mo) |
License | Proprietary CLI | MIT, full source |
Lock-in | Anthropic cloud + enterprise contract | BYOK, open adapters, remote runtime |
Both harnesses ran the same Kimi K3 model in the FrontierHarness comparison. MiniMax's 76.7% / $1.82 result is MiniMax's own run of the same task suite, not an entry on the independent leaderboard.
🔍 THE AUDIT
5 CRITICAL QUESTIONS, OCTOBER EDITION
Q1: What does a completed task actually cost?
A: On the independent board, Claude Code is the cost outlier at $18.34 per completed task. MiniMax's own run reports $1.82. At 1,000 completed tasks a month, that is $198,240 a year in difference. Even if you discount the vendor number by 50%, the gap still pays for a senior engineer.
Q2: Is MiniMax's number independent?
A: No. FrontierHarness v1.0 tested 12 harness configurations on 30 tasks — 21 Terminal-Bench 2.1 plus 9 DeepSWE v1.1 — for 360 trials. MiniMax Code is absent from the public board. Claude Code's 63.3% and $18.34 are third-party measurements. Keep that asymmetry in mind when someone forwards a launch deck.
Q3: What changed in October?
A: Claude Code v2.1.296 made auto mode the default and moved subagents to the background. MiniMax 0.6.12 added retired-model auto-routing, a standalone App Server for SSH remote runtimes, plugin-declared subagents, and smarter compaction. Migration clock: M2.5 and M2.7 stopped accepting new users on October 19 and shut down fully on November 18.
Q4: Which model should sit under the harness?
A: M3.1 Flash for cheap volume: $0.15 / $0.60, 1M context, 200+ TPS, 73.7% on DeepSWE v1.1. Kimi K3.1 Pro for hard tasks: 65.7% on Terminal-Bench 4.0 against Opus 5.5's 66.4%, at roughly $1.8 / $9.6 per million tokens. DeepSeek V4.1 Flash for overnight batch: $0.15 / $0.60 off-peak, half that again on cache hits.
Q5: What is the real lock-in risk?
A: Claude Code is a proprietary CLI tied to Anthropic's cloud. MiniMax Code is MIT-licensed, supports BYOK and open adapters, and now runs a remote runtime over SSH. The trade-off is churn: 13 releases in 15 days means pin your version, test before you roll, and keep a rollback image.
📡 ECOSYSTEM RADAR
Your China Monday Briefing
China Insights Weekly is a free Monday briefing on the AI, EV, robotics, trade and consumer stories coming out of China, and why they matter for business everywhere.
Written by Yereth Jansen and Tomas Kucera, with some 30 years in China between them.
📈 BENCHMARK
WHAT THE NUMBERS ACTUALLY SAY
1. The independent board: 17.5x spread, same model
Source: MiniMax, September 2026. Vendor self-report.
FrontierHarness fixed Kimi K3, used identical infrastructure, and varied only the harness. Pass rates landed between 50.0% and 66.7%. Cost per completed task landed between $1.05 and $18.34.
Quality is roughly comparable. The invoice is not. That is the whole argument for treating the harness as an architectural decision, not a developer preference.
Claude Code posted 63.3% at $18.34 and 9m 38s. Codex posted 66.7% at $3.47 and 6m 43s. Pi posted 60.0% at $2.43. The cheapest configuration, Exo Harness, posted 53.3% at $1.05.
2. Unit cost per successful task
Source: MiniMax, September 2026. Vendor self-report.
$18.34 vs $1.82 is a 10.1x gap on the same model. Your team is not paying for intelligence. It is paying for context management.
At 100 tasks a month the difference is $1,652. At 1,000 it is $16,520 a month. At 10,000 it is a hiring decision.
3. The October price sheet
Model | Input / output per 1M | Context | Why it matters |
|---|---|---|---|
Claude Opus 5.5 | $4 / $20 | 1M | Default on every paid Claude Code plan |
Claude Sonnet 5.5 | $2 / $10 | — | The cheaper Claude route for routine work |
Claude Haiku 5.5 | $0.10 / $0.50 | — | Prompts up to 100k tokens; $0.50 / $2.50 above |
Kimi K3.1 Pro | ~$1.8 / ~$9.6 | 1M | 65.7% on Terminal-Bench 4.0; weights due Nov 4 |
MiniMax M3.1 Flash | $0.15 / $0.60 | 1M | 200+ TPS; 73.7% on DeepSWE v1.1 |
DeepSeek V4.1 Flash | $0.15 / $0.60 | 1M | Off-peak; peak is 2x; cache hits $0.003 |
Prices checked October 25, 2026. Chinese-lab prices converted at roughly 7.1 CNY/USD; FX moves the last digit, not the conclusion.
⚖️ VERDICT
WHICH STACK BELONGS ON YOUR BALANCE SHEET?
Choose Claude Code if: your work is dominated by irreversible, high-context engineering — 600k-line migrations, gnarly architecture changes, security-sensitive refactors — and you already have an Anthropic enterprise agreement with SSO and audit trails. Opus 5.5 is genuinely strong. Budget $18 per task and stop pretending it is a seat cost.
Choose MiniMax Code if: your workload is high-volume and reversible — test-fix loops, dependency upgrades, CI triage, repetitive refactors. The MIT license, BYOK adapters, and $1.82 vendor-reported task cost are a serious combination. Accept the 0.6.x churn and verify the benchmark on your own repos.
ShanghAI Choice — the two-lane stack: Claude Code + Opus 5.5 for the 10% of tasks where a wrong move is expensive. MiniMax Code 0.6.12 (or Pi/OpenCode) with M3.1 Flash or Kimi K3.1 Pro for the 90% that are reversible. Route by task class, not vendor loyalty. Expected blended saving: 60% to 80% without touching headcount.
THIS WEEK OFFER (Affiliate): Notion Business — 3 MONTHS FREE
📋 APPLY TO YOUR BUSINESS
A 4-STEP OCTOBER MIGRATION PLAYBOOK
This is the non-technical version. Four steps, two weeks, one decision your finance team can defend.
Step 1: Measure cost per task, not cost per seat.
Pull 30 days of agent sessions. Tag them by task type. Divide total token spend plus retries by merged pull requests. If you cannot produce that number, you are buying on vibes.
Step 2: Split the work into two lanes.
Lane A is irreversible: architecture, migrations, security. Lane B is reversible: tests, dependency bumps, lint fixes, CI triage. Set a hard rule: any task projected above $5 gets Lane A treatment.
Step 3: Run a 20-task pilot on Lane B.
Install MiniMax Code 0.6.12, pin the version, enable permission gates, and run M3.1 Flash first. Then rerun the same 20 tasks with Kimi K3.1 Pro via BYOK. Compare pass rate and time-to-merge against your Claude Code baseline.
Step 4: Audit the invoice, then commit.
Require an independent harness run, not a vendor PDF. If Lane B clears 80% of your Claude pass rate at one-fifth the cost, move it permanently and renegotiate your Claude seat count. If it does not, you have lost two weeks and a few hundred dollars — cheap tuition.
SCALE THIS ARCHITECTURE
If your leadership team wants to audit enterprise AI privacy, escape vendor lock-in, or build a cost-effective European sovereign AI stack:
We review your active pipelines, cut 50% to 70% of redundant compute and subscription overhead, and implement high-efficiency cloud runtimes directly into your stack.
Until next week,
The ShanghAI Guy
