GPT-6.2 Astra and Sol: Cheaper, Not Smarter
Official benchmarks, flat AA Coding Agent scores, and a 59% hallucination rate on GPT-6.2 Astra.
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.
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:
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.nextDelayMs cap: ensure exponential backoff plus random jitter never exceeds maxMs and stays non-negative.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.
ompBoth models produced clean code that passed all 7 tests. The difference was how long they took and how many tokens they burned getting there.
| Metric | Claude Sonnet 5.5 | GPT-6.2 Astra | Astra / Sonnet 5.5 |
|---|---|---|---|
| Test suite | 7 pass, 0 fail | 7 pass, 0 fail | Tie |
| Wall-clock time | 23.96s | 249.57s | 10.4x |
| Turns | 3 | 10 | 3.3x |
| Tool calls | 2 (bash x2) | 13 (todo x4, read x4, glob, edit x2, bash x2) | 6.5x |
| Input tokens | 6 | 35,179 | — |
| Output tokens | 2,071 | 5,547 | 2.7x |
| Cache read tokens | 44,663 | 137,216 | 3.1x |
| Cache write tokens | 9,890 | 0 | — |
| Total tokens | 56,630 | 177,942 | 3.1x |
| Official API cost | $0.0499 | $0.5582 | 11.2x |
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.
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:
| Model | Input | Output | Cache read | Cache 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:
| Monthly volume | Claude Sonnet 5.5 | GPT-6.2 Astra | Extra 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.
Our omp run matches what independent coding leaderboards already show:
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).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).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 | Recommended model | Why |
|---|---|---|
| Daily bug fixes, backend logic, test loops | Claude Sonnet 5.5 | 10x 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 work | Claude Sonnet 5.5 (xHigh) | Ranks 3rd on CodeArena WebDev (1,772 Elo) with predictable diffs. |
| Scientific terminal tasks, 3D, deep code review | GPT-6.2 Astra or Sol | Astra 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.
CodingPass Go includes $60 of Sonnet 5.5 a month for $10.Metered at official API rates, cancel anytime.