查看中文版 →

Deutsche Version ansehen →

Открыть русскую версию →

Ver versión en español →

Log inChoose a plan
Blog · Model tests

GPT-6.2 Astra Still Loses to Sonnet 5.5

OpenAI pitched GPT-6.2 Astra as a cheaper flagship at $8 / $40 per million tokens. On an identical two-file Bun TypeScript bug fix in omp, Claude Sonnet 5.5 finished in 23.96s for $0.0499, while GPT-6.2 Astra burned 4m 10s and $0.5582—10.4x slower and 11.2x the bill.

OpenAI framed the GPT-6.2 Astra launch around one metric: lower cost per task. List prices dropped 20% from GPT-6 Astra's $10 / $50 per million tokens down to $8 / $40, and OpenAI highlighted a $2.89 per-task bill on Agents' Last Exam.

Being cheaper than your own predecessor is not the same as beating the mid-tier workhorse developers actually use. Even during the GPT-6 Astra cycle, Anthropic's Claude Sonnet 5.5 already undercut Astra on Artificial Analysis's cost-per-task charts. With GPT-6.2 Astra, OpenAI's flagship still loses to Sonnet 5.5 on everyday coding speed and cost.

Bottom line: On November 19, we ran both models through omp on the same Bun + TypeScript bug-fix and implementation task. claude-sonnet-5-5 passed all 7 tests in 23.96s across 3 turns for $0.0499 at official API rates. gpt-6.2-astra took 249.57s, 10 turns, 13 tool calls, and $0.5582—10.4x slower and 11.2x more expensive for the exact same passing diff.

The Test: A Routine Backend Bug-Fix Task

Rather than relying only on aggregate leaderboards, we set up a clean, isolated three-file Bun TypeScript project (webhook-kit) and gave both models the exact same environment (omp v19.0.0, --mode json -p --no-session) and prompt, ending only when bun test passed.

The repo started with 2 passing tests and 5 failures across three requirements:

  • Fix verifyWebhookSignature in src/webhook.ts: handle comma-separated headers with whitespace (t=..., v1=deadbeef, v1=...), support multiple v1 signatures, reject future timestamps outside tolerance via Math.abs(now - timestamp) > tolerance, and replace === with timingSafeEqual.
  • Fix nextDelayMs cap: ensure exponential backoff plus random jitter never exceeds maxMs and stays non-negative.
  • Implement IdempotencyStore in src/idempotency.ts: cache resolved values for ttlMs from completion time, deduplicate concurrent in-flight calls for the same key, and never cache rejected promises.

There are no obscure frameworks or sprawling context traps here—just the kind of routine two-file change developers run dozens of times a day.

Head-to-Head Results in omp

Both models produced clean code that passed all 7 tests. The difference was how long they took and how many tokens they burned getting there.

Head-to-head omp benchmark: Claude Sonnet 5.5 vs GPT-6.2 Astra
MetricClaude Sonnet 5.5GPT-6.2 AstraAstra / Sonnet 5.5
Test suite7 pass, 0 fail7 pass, 0 failTie
Wall-clock time23.96s249.57s10.4x
Turns3103.3x
Tool calls2 (bash x2)13 (todo x4, read x4, glob, edit x2, bash x2)6.5x
Input tokens635,179—
Output tokens2,0715,5472.7x
Cache read tokens44,663137,2163.1x
Cache write tokens9,8900—
Total tokens56,630177,9423.1x
Official API cost$0.0499$0.558211.2x

The tool trace explains the gap. claude-sonnet-5-5 inspected the files and applied both fixes in one pass, ran bun test on the next turn, and summarized in 24 seconds.

gpt-6.2-astra treated a three-file directory like a multi-week migration: it called todo four times, glob once, and read four separate times before finishing its edits, dragging a sub-minute task out to 4 minutes and 10 seconds while generating 2.7x as many output tokens.

Why the Bill Explodes: 4x Unit Price Times 3x Volume

At official API list rates per million tokens, GPT-6.2 Astra already costs 4x as much as Claude Sonnet 5.5 in every token category:

Official API list pricing per million tokens
ModelInputOutputCache readCache write
Claude Sonnet 5.5 (claude-sonnet-5-5)$2.00$10.00$0.10$2.50
GPT-6.2 Astra (gpt-6.2-astra)$8.00$40.00$0.40$10.00

Multiply a 4x higher rate card by 3.1x total token volume across 10 turns, and the per-task bill jumps to 11.2x. Scaled across a team's monthly workload, the difference adds up quickly:

Projected monthly workload cost based on measured omp run
Monthly volumeClaude Sonnet 5.5GPT-6.2 AstraExtra spend
1,000 similar tasks$49.90$558.20+$508.30
10,000 similar tasks$499.00$5,582.00+$5,083.00

GPT-6.2 Astra also carries a steep long-context cliff: once a single request crosses 272K input tokens, input and cache-read rates double and output jumps 1.5x for the entire request.

Public Benchmarks Tell the Same Story

Our omp run matches what independent coding leaderboards already show:

  • Artificial Analysis Coding Agent Index: GPT-6.2 Astra scored 53, unchanged from GPT-6 Astra (53), and five points behind Opus 5.5 and Fable 5.2 in Claude Code (58).
  • CodeArena WebDev (Nov 14, 2026, 159 models, >1M votes): Claude Sonnet 5.5 (xHigh) ranks 3rd globally at 1,772 Elo, ahead of GPT-6 Astra (1,763, 5th) and GPT-6.1 Sol (1,742).
  • Artificial Analysis Intelligence Index task cost: Even before the 6.2 release, Sonnet 5.5 at low ($0.41), medium ($0.54), high ($1.08), and max ($2.74) all cost less per task than GPT-6 Astra (high) ($3.26).
  • Higher hallucination rate: On AA-Omniscience, GPT-6.2 Astra's hallucination rate climbed from 50% (GPT-6 Astra) to 59%.

When to Use Each Model

GPT-6.2 Astra has real strengths: it leads Terminal-Bench-Science 0.1 at 77.1%, tops AA-3D at 1,318 Elo, and catches 96% of bugs in xhigh code-review runs.

For everyday software engineering—fixing bugs, writing endpoints, building UI, and running test loops—routing a 24-second task to a 4-minute flagship wastes both developer time and budget.

Workload guide: Claude Sonnet 5.5 vs GPT-6.2 Astra
WorkloadRecommended modelWhy
Daily bug fixes, backend logic, test loopsClaude Sonnet 5.510x faster in our omp test, 1/4 the list price, and ~$0.05/task. CodingPass Go ($10/mo, $60 allowance) covers ~1,200 similar runs.
Frontend & UI workClaude Sonnet 5.5 (xHigh)Ranks 3rd on CodeArena WebDev (1,772 Elo) with predictable diffs.
Scientific terminal tasks, 3D, deep code reviewGPT-6.2 Astra or SolAstra leads science and 3D benchmarks; for pure code review, GPT-6.2 Sol ($2 / $10) gives most of the value at a quarter of Astra's price.

If your daily workflow is shipping code, you do not need a Pro-tier flagship. CodingPass Go ($10/month) includes claude-sonnet-5-5 with a $60 monthly allowance—save Astra for the rare scientific or 3D task that actually needs it.

Boarding

Stick to Sonnet 5.5 for daily coding.

CodingPass Go includes $60 of Sonnet 5.5 a month for $10.Metered at official API rates, cancel anytime.

  • Request ID on every response
  • Full refund if you've used 5% or less
  • No surprise charges
Choose GoClaude Code Guide

$10/mo until you cancel.