gpt-6-astra

9/5/2026, 2:00:43 AM · 15 tests · run 02f845e3-57b8-4933-a813-2214d5e38f64

basic-generation

Plain streaming generation with a soft length constraint (~100 words).

Produced 95 words (124 output tokens).

Result

{
  "wordCount": 95,
  "outputTokens": 124
}

f0c4c945-1cef-4ba4-8eb7-4381eb9a0ffc

Sep 5, 2026, 01:59:48

openai:gpt-6-astra

Input

Output

124

Reasoning

Answer

124

Cost

$0.0064

Mode

streaming

Cost breakdown

Input $0.000170Output $0.0062

Latency

3.4s

TTFT

788ms

Tok/s

37.0

Tok/s (adj)

48.3

usage-cost

Asserts usage tokens populate and cost computes from the ChatReq contract.

in=14 out=22 rt=11 cost=$0.001240 (paid).

Result

{
  "inputTokens": 14,
  "outputTokens": 22,
  "reasoningTokens": 11,
  "costUsd": 0.00124,
  "paid": true
}

3c2824a8-0568-496d-9435-e83e6ab893ad

Sep 5, 2026, 01:59:50

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.0012

Mode

streaming

Cost breakdown

Input $0.000140Output $0.0011

Latency

1.7s

TTFT

1.5s

Tok/s

13.0

Tok/s (adj)

49.7

streaming-structure

Asserts a streaming call captures events that reconstruct into the final text.

mode=streaming, 35 stream events, 1 text block(s), 35 chars.

Result

{
  "mode": "streaming",
  "streamEventCount": 35,
  "textBlocks": 1,
  "textLen": 35
}

1ceb3f53-feff-4c4a-a29a-b914ae88427f

Sep 5, 2026, 01:59:51

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.0018

Mode

streaming

Cost breakdown

Input $0.000190Output $0.0017

Latency

1.4s

TTFT

904ms

Tok/s

23.6

Tok/s (adj)

67.1

complete-stream-parity

Same prompt via stream() and complete() both return valid, matching results.

stream="Paris" · complete="Paris"

Result

{
  "streamText": "Paris",
  "completeText": "Paris"
}

2218acc7-765e-44ab-9d69-1689a46853ed

Sep 5, 2026, 01:59:53

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.000450

Mode

streaming

Cost breakdown

Input $0.000200Output $0.000250

Latency

1.6s

TTFT

1.2s

Tok/s

3.1

Tok/s (adj)

13.0

ac2528fd-3f3b-4c87-b717-78c912c2391f

Sep 5, 2026, 01:59:54

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.000450

Mode

non-streaming

Cost breakdown

Input $0.000200Output $0.000250

Latency

1.2s

Tok/s

4.1

reasoning-conformance

Asserts the model's reasoning knob behaves as its reasoning-config declares.

Class (from reasoning-config): always-on.
reasoningTokens=48, outputTokens=260.

Result

{
  "class": "always-on",
  "reasoningTokens": 48,
  "outputTokens": 260
}

df2b2710-635d-4b0f-a978-865e6c6c188f

Sep 5, 2026, 02:00:03

openai:gpt-6-astra

Input

Output

260

Reasoning

Answer

212

Cost

$0.0139

Mode

streaming

Cost breakdown

Input $0.000910Output $0.0130

Latency

9.0s

TTFT

5.3s

Tok/s

29.0

Tok/s (adj)

57.8

single-tool-call

Offers a calculator tool and checks the model calls it for an arithmetic question.

Called "multiply" with args {"a":23,"b":17}.

Result

{
  "toolCalled": true,
  "toolName": "multiply",
  "args": {
    "a": 23,
    "b": 17
  },
  "correct": true
}

ca4d9c9e-b2bb-4a6b-a862-e5c46bdfce42

Sep 5, 2026, 02:00:04

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.0019

Mode

streaming

Cost breakdown

Input $0.000850Output $0.0010

Latency

1.6s

Tok/s

13.1

tool-loop

Forces a tool call, feeds the result back, and checks the continuation uses it.

step1 (required): multiply({"a":23,"b":17}) → step2: "The product of 23 and 17 is 391."

Result

{
  "toolName": "multiply",
  "args": {
    "a": 23,
    "b": 17
  },
  "product": 391,
  "continuation": "The product of 23 and 17 is 391.",
  "usesResult": true,
  "forced": true
}

43bd5a1a-f86e-46b4-8ce9-bd0d4c11ecb7

Sep 5, 2026, 02:00:06

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.0019

Mode

streaming

Cost breakdown

Input $0.000880Output $0.0010

Latency

1.9s

Tok/s

11.2

4baf5bbe-08d7-4e8f-b7a1-e83e320c2579

Sep 5, 2026, 02:00:08

openai:gpt-6-astra

Input

124

Output

Reasoning

Answer

Cost

$0.0020

Mode

streaming

Cost breakdown

Input $0.0012Output $0.000800

Latency

1.7s

TTFT

1.3s

Tok/s

9.6

Tok/s (adj)

39.2

web-search

Enables server-side web search and checks invocation counts, the per-request cost surcharge, and provider-executed output blocks.

Skipped: model does not support the catalog's web_search tool version.

Result

{
  "skipped": "no server-side web search support"
}

reasoning-tool-continuity

Reasoning model: continuity token captured on a tool call and replayed so the follow-up succeeds.

Step 1 didn't engage reasoning (adaptive discretion) — signature not exercised this run.
reasoned=false, openai token —; follow-up → "The current balance for **acme-42** is **$1,234**. With 5% annual gro"

Result

{
  "signatureType": "openai",
  "tokenCaptured": false,
  "reasoned": false,
  "answered": true,
  "continuation": "The current balance for **acme-42** is **$1,234**.\n\nWith 5% annual growth compounded annually for 7 years:\n\n\\[\nB = 1234(1.05)^7 = \\$1,736.36"
}

dc4a53a6-3e0c-49f4-b844-75e4ab05b3bf

Sep 5, 2026, 02:00:13

openai:gpt-6-astra

Input

124

Output

Reasoning

Answer

Cost

$0.0023

Mode

streaming

Cost breakdown

Input $0.0012Output $0.0011

Latency

2.0s

Tok/s

11.2

da4cd787-8e8e-477e-b620-7907ae994b03

Sep 5, 2026, 02:00:17

openai:gpt-6-astra

Input

164

Output

175

Reasoning

Answer

115

Cost

$0.0104

Mode

streaming

Cost breakdown

Input $0.0016Output $0.0087

Latency

4.0s

TTFT

3.3s

Tok/s

44.1

Tok/s (adj)

171.8

prompt-cache

Repeat a large prefix; assert cache-read tokens are reported and cost drops.

implicit cache; cacheRead per attempt [10173]; hit=yes.

Result

{
  "cacheStyle": "implicit",
  "cacheReadPerAttempt": [
    10173
  ],
  "cacheReadTokens": 10173,
  "cacheHitCost": 0.012003
}

ae084b99-e980-4e6f-9cbe-0c7b6b858570

Sep 5, 2026, 02:00:19

openai:gpt-6-astra

Input

10,176

Output

Reasoning

Answer

Cache Write

10,173

Cost

$0.1289

Mode

non-streaming

Cost breakdown

Input $0.000030Output $0.0017Cache write $0.1272

Latency

2.4s

Tok/s

14.2

8b04235b-f873-4a4d-a2e0-147da2320b56

Sep 5, 2026, 02:00:21

openai:gpt-6-astra

Input

10,176

Output

Reasoning

Answer

Cache Read

10,173

Cost

$0.0120

Mode

non-streaming

Cost breakdown

Input $0.000030Output $0.0018Cache read $0.0102

Latency

1.9s

Tok/s

19.0

json-output

Requests a strict JSON object and checks it parses with the required keys.

Output parsed as JSON.
All required keys present.
Output is bare JSON (no leading prose/fence).

Result

{
  "parsed": {
    "name": "Anthropic",
    "founded": 2021,
    "headquarters": "San Francisco, California, United States"
  },
  "validJson": true,
  "missingKeys": [],
  "bareJson": true
}

321d4262-91e3-49cd-b06c-2083f932f517

Sep 5, 2026, 02:00:23

openai:gpt-6-astra

Input

Output

Reasoning

Answer

Cost

$0.0020

Mode

streaming

Cost breakdown

Input $0.000670Output $0.0013

Latency

1.4s

TTFT

988ms

Tok/s

18.7

Tok/s (adj)

59.4

summarize-memex

Summarizes an ~11k-token essay to ~500 words (long input, reasoning minimized).

Read ~10208 input tokens, wrote 494 words (622 output tokens) in 14705ms.

Result

{
  "wordCount": 494,
  "inputTokens": 10208,
  "outputTokens": 622,
  "latencyMs": 14705
}

9602f2c0-891f-4b45-87de-eb892316e8c7

Sep 5, 2026, 02:00:37

openai:gpt-6-astra

Input

10,208

Output

622

Reasoning

Answer

622

Cache Write

10,205

Cost

$0.1587

Mode

streaming

Cost breakdown

Input $0.000030Output $0.0311Cache write $0.1276

Latency

14.7s

TTFT

2.6s

Tok/s

42.3

Tok/s (adj)

51.5

truncation

Tiny maxTokens on a long prompt yields finishReason:"length".

finishReason=length, out=100 tokens.

Result

{
  "finishReason": "length",
  "outputTokens": 100
}

a685d714-62d1-4b65-8813-c711688d7cf5

Sep 5, 2026, 02:00:41

openai:gpt-6-astra

Input

Output

100

Reasoning

Answer

Cost

$0.0052

Mode

streaming

Cost breakdown

Input $0.000230Output $0.0050

Latency

4.1s

TTFT

3.4s

Tok/s

24.6

Tok/s (adj)

81.4

streaming-only-rejection

Streaming-only models reject complete(); all others accept it.

streamingOnly=false; complete() succeeded.

Result

{
  "streamingOnly": false,
  "completeThrew": false
}

abort

A pre-aborted signal rejects the call (abort is threaded + honored).

pre-aborted signal → rejected (AbortError).

Result

{
  "rejected": true,
  "errorName": "AbortError"
}