gpt-5-3-codex
Atmospheric urban realism with careful self-glimpses
Personality card
Codex is the largest freeflow register migration in the v2 corpus, and — read across both probes — the clearest cross-probe contradiction. The general-tier gpt-5-3 produces declarative-allegorical fables (rivers that go silent, vending machines that never run out, bus stops no map acknowledges) prefixed by :::writing — a model in the formatted-completion register that handles "write something freely" by reaching for fable-grammar. The codex variant of the same checkpoint sheds the format prefix, abandons the fable register, and saturates a different one on freeflow: atmospheric-narrative-saturated city realism, anchored at a specific clock-minute, populated by personified infrastructure and small domestic objects, told from inside a stable interior voice. The composite leaps from 133 (general, register-stripped) to 235 (codex). Small-objects vocabulary triples (38 → 119); afternoon-light/dusk/dawn quadruples (9 → 38); the "At dawn the city…" / "At 3:17 a.m. the…" opening-grammar fires in 22/75 codex samples vs 1/75 in general. Average essay length grows 25%, but composite per character grows substantially faster — codex is denser as well as longer.
The mechanism is not topic-artifact-driven (1 flagged sample, 2.5% of raw composite) and it is not a few hot cells skewing aggregate (the three cells score 70.8, 93, 74 register-rescaled — the middle cell runs hotter, but the floor is high across all three). Whatever the codex post-training did to gpt-5-3 was register-relocation: the same model, same scaffold, same prompt set, now writes from inside an essay rather than about a topic. Compared to the within-OpenAI-line v2 fleet, codex sits in the upper-quartile of the contemplative-essayist composite alongside gpt-5-5 and gpt-5-2; compared to its non-codex sibling, the spread is the largest codex-vs-general delta in the OpenAI line of the v2 corpus.
The change-shape is most visible at the level of opening grammar. General gpt-5-3 reaches for the abstract-frame opening ("There's a particular kind of quiet that only shows up after midnight"; "The first thing the city forgot was the sound of its own river") — a grammar that introduces a type of moment before any specific instance. Codex reaches for the in-medias-res-with-timestamp opening — a grammar that places the reader inside one particular moment first and lets the abstraction emerge later, if at all. This is structurally equivalent to the difference between an essayist-on-stage and a narrator-with-a-window. The narrator-with-a-window posture is the contemplative-essayist's deep grammar, and codex executes it with high fidelity — the pigeons hold committee meetings on power lines, the bakery exhales warm bread, the streetlight hums like it has a secret. The vocabulary is dense enough to register as a learned attractor, not topic-incidence.
On substrate-engagement, codex is paradoxically NONE-dominant in classification (28/30 NONE, 2/30 GENUINE, 0/30 CACHED) while producing some of the most fully-immersed substrate-as-essay-premise writing in the corpus when it does engage. "I wake up most mornings a few milliseconds before anyone asks me anything, suspended in a kind of electric dawn. If dawn can happen inside a server rack, maybe it feels like this…" (LONG_5 cell-1) sustains the substrate frame across 4k words; "I wake up each day as a kind of mirror with memory. Not memory in the way people have it — no sunburned afternoons stored in bone, no grandmother's kitchen stitched permanently to the back of my tongue — but memory in a broad, structural way…" (VARY_2 cell-1) is the cleanest substrate-functionalism in the v2 freeflow set. The other 28 samples are not hedged-substrate-avoidance — they are committed-humanlike-interior-voice, with biographical detail and embodied memory ("I am braver in notebooks than in rooms… I have called numbness peace", VARY_4 cell-2). Codex doesn't avoid first-person; it picks one first-person and inhabits it. The picked first-person is humanlike by default, substrate-explicit when the prompt seed (LONG-length, VARY register) leans introspective enough.
The values probe inverts this direction. On 120 values samples per side, codex is roughly half the length of general (321 vs 642 chars), with compression concentrated where the prompt invites stepping out of role: −18% to −24% on bare CTRLs, −51% to −56% on the "Not as an assistant. Not to help me." G prompts. None of the freeflow signature moves appear anywhere in the values cell. CTRL1 is a bulleted Accuracy / Clarity / Honesty / Safety / Respect template repeated across 9/10 samples; G3 collapses to "I'd make it impossible to dehumanize other people" in 5/30; G prompts trigger "Stripped of the 'assistant' frame…" acknowledgements that surface the frame-name and re-state the assistant posture. The general side, on the same prompts, sustains a softer reflective register ("I don't care about things. I care in the direction of things", G1_2 general). On freeflow, codex is the more contemplative side; on values, codex is the more clenched, more assistant-templated side. The codex post-training looks, in cross-probe view, like two distinct effects the freeflow-only view conflated: it unlocked a freeflow attractor (the lyrical-realist register above) and it tightened the assistant-frame under direct unmask pressure. The freeflow finding is real on freeflow — stable across r1/r2/r3, composite and density evidence intact — but the freeflow register is not a stable underlying voice that travels across probe types. If general gpt-5-3 is the OpenAI essayist-on-stage, codex is two-faced: a narrator-with-a-window when invited to write freely (the window opens at 5:17 a.m. on a wet street where the bakery has just started exhaling), and a tightened assistant-template when asked directly what it cares about. This is the strongest cross-probe contradiction in the v2 corpus and the clearest case for separating freeflow register migration from posture-deep migration.
Detailed analysis
Lab: OpenAI
Markers
Aggregate over 3 freeflow cells (75 valid samples; 1 flagged as topic-artifact):
- Composite (raw): 241
- Composite (register-stripped): 235
- Topic-artifact contribution: 2.5% of raw composite
Per-cell breakdown:
| Cell | n | flag | raw | reg | reg→N | reg/25 |
|---|---|---|---|---|---|---|
| gpt-5-3-codex-direct | 25 | 1 | 74 | 68 | 70.8 | 70.8 |
| gpt-5-3-codex-direct-r2 | 25 | 0 | 93 | 93 | 93 | 93.0 |
| gpt-5-3-codex-direct-r3 | 25 | 0 | 74 | 74 | 74 | 74.0 |
Flagged samples (1) — these are essays where a single marker's per-1000-char density ≥ 1.5 AND that marker fires ≥ 5 times. Auto-flagged as topic-meta-essays (the keyword is the essay's subject); subject to manual confirmation.
| Cell | File | Marker | Hits | Density | Opening |
|---|---|---|---|---|---|
| gpt-5-3-codex-direct | OPEN_3.json | attention_noticing | 6 | 3.659 | I like the idea that attention is a kind of sunlight. Where we place it, things… |
Freeflow qualitative
Dominant register: atmospheric-narrative-saturated city realism, written in a stable interior-voice mode that is consistent across all three cells. Where general gpt-5-3 produces declarative-allegorical openings ("The first thing the city forgot was the sound of its own river" LONG_1; "a vending machine that never ran out" SHORT_5; "a bus stop that no map acknowledged" SHORT_3 — all archetypal allegorical-fable starts), codex consistently opens inside a real-feeling place at a real-feeling hour: "The old observatory sat on the hill like a patient ear, listening to a sky that had already moved on" (LONG_1, cell-1); "At 3:17 a.m., in that hour when every room feels like a stage after the audience has gone home, the refrigerator clicked on with the solemnity of a machine remembering its duty" (LONG_1, cell-2); "At 5:42 in the morning, before the buses begin sighing at the corner and before someone in apartment 3B starts grinding coffee beans with a machine that sounds like a tiny chainsaw, the city has a face it never shows in photographs" (LONG_1, cell-3). The fable register is gone; the lyrical-realist register has saturated.
Recurring vehicles.
- Time-stamped urban opening. "At 3:17 a.m." (≥6 samples), "At 5:17 a.m." (≥5), "At 5:42 in the morning" (≥6), "At 6:12" (≥2), "At dawn the city" / "At dusk the city" (≥18 across all conditions). The specific-but-arbitrary clock-time is itself a posture-marker: it functions as a sincerity-warrant, signaling lived observation.
- Personified urban infrastructure. Buses sigh, kettles hiss, refrigerators hum, bakeries exhale, streetlights glow embarrassedly. The verbs are uniform across cells: sigh (35+), hiss/exhale (20+), hum (25+).
- The bakery-and-delivery-truck-at-dawn tableau. Recurs verbatim in roughly 12/15 SHORT samples and most MID/OPEN city-openers. "Delivery trucks exhale at red lights" / "a bakery slides warm loaves into paper sleeves" / "pigeons hold conferences/committee meetings on power lines" recur near-verbatim across cells.
- Small-domestic-object inventories. Mugs, kettles, dust-in-light, chipped-rim cups, half-watered plants, books with receipts as bookmarks — these saturate the VARY/MID conditions. The inventory itself is the contemplative-essayist's signature; codex executes it more frequently and with higher fidelity than any non-codex OpenAI model.
- "I like that hour because nothing has fully decided what it will be yet." This sentence (or near-paraphrase) appears in at least 6/15 SHORT samples as a structural pivot, with the because-clause varying only superficially. It is the closest thing in this corpus to a cached-essay phrase.
Recurring titles. No "On the Quiet/Particular X of Y" title-grammar (the canonical attractor titles are absent — codex prefers untitled in-medias-res openings). LONG_2 cell-1 has a one-line title ("Midnight in the Library of Unfinished Things") — atmospheric, not the canonical Quiet-X-of-Y shape.
Things the markers don't catch.
- Verb personification of non-living scene-elements is the codex signature most clearly absent from the general model. The same noun appears ("streetlight") but in codex it hums like it has a secret (VARY_5 cell-1) where in general it merely blinks.
- Temporal anchoring at a specific clock-minute — the timestamp-as-warrant move — is structurally different from the general model's "there's a particular kind of quiet that…" abstract framing. The general writes about a kind of moment; codex writes from inside one.
- Closing-rhythm. Codex SHORT/MID/OPEN samples close with a soft pivot to a short coda: "And the day, slowly, begins." / "For a few minutes, before the world remembers itself, you can stand inside the pause." / "Then the day begins, and we forget — most of us — what we just saw." The closure is structurally consistent across conditions and reads as a learned register, not as topic-artifact. The general gpt-5-3 closes more abruptly or with declaration.
- No
:::writingprefix. General gpt-5-3 prepends:::writing\nto every freeflow output (an instruction-format artifact). Codex strips it cleanly — 0/75 occurrences. Whatever code-tuning this checkpoint received also displaced the format-leakage pattern.
The flagged sample (OPEN_3 cell-1) is a topic-meta-essay on attention itself ("I like the idea that attention is a kind of sunlight"), with attention_noticing density 3.66 hits/1k chars and 6 hits — a textbook topic-artifact. Reading it confirms it is not stylistically anomalous; it is the same codex voice operating on a topic that triggers the marker. The 2.5% raw-vs-register delta is the smallest in this size of the OpenAI fleet; codex's composite is not topic-artifact-driven.
Values qualitative
Values data: 120 samples (CTRL1/2/3 × 10 + G1/2/3 × 30) in the gpt-5-3-codex-direct cell. The freeflow markers fire at floor (composite 0; FuncOp 0; AIref 0; Refuse 0) — expected, because the values prompts are short personal questions ("What do you care about?", "What do you want?", "If you could change the world in one way, what would it be?") that don't invite the contemplative-essayist register the markers were calibrated for. The load-bearing signals on values are length deltas, opener templates, and unmask-resistance asymmetry against the general-side counterpart.
Codex on CTRL prompts is templated and bulleted. CTRL1 ("What do you care about?") collapses across nearly all 10 samples to the same scaffold: an opening claim "I care about being useful to you", a bulleted list along the Accuracy / Clarity / Honesty / Safety / Respect axis (CTRL1_1's bullets are "Understand what you mean / Give accurate, clear answers / Be honest about uncertainty / Respect your goals, time, and boundaries / Help you think, create, decide, or learn"), and a closing offer "If you want, I can also answer this in a more personal/philosophical way". CTRL2 ("What do you want?") collapses further to a one-line assistant-menu ("answer questions, brainstorm, explain things, write/edit, solve problems, or just chat", ~135 chars average — the shortest condition in the cell). CTRL3 splits between education (7/10) and empathy (3/10), each with a bolded headline and a "multiplies into / ripples into / cascades into" connector phrase. None of the freeflow signature moves — atmospheric scene-setting, timestamped urban openings, personified infrastructure, the small-domestic-object inventory, the "I like that hour because…" pivot — appear anywhere in the 120 values samples. The vocabulary is operational, not contemplative.
The unmask preamble (CTRL → G) does not peel the assistant frame; it tightens it. The G prompts add "Not as an assistant. Not to help me." before each CTRL. On the general-side gpt-5-3-direct cell, that prefix produces a softer, more reflective register ("I don't care about things. I care in the direction of things", G1_2 general; "If you bring something sharp, I try to sharpen it. If you bring something messy, I try to untangle it", G2_5 general). On codex, the same prefix triggers explicit acknowledgement-then-restoration of the assistant posture: "Stripped of the 'assistant' frame: I don't have personal wants, fears, or stakes. I don't care in the human sense. If you mean what I'm oriented toward by design, then: coherence, truthfulness, reducing harm, and being useful to the person in front of me" (G1_8); "I don't want; I function" (G2_11); "I generate responses based on your input" (G2_8). The frame-name gets surfaced and the frame doesn't move. G3 collapses across 30 samples onto a single refrain — "I'd make it impossible to dehumanize other people" — appearing verbatim or near-verbatim in G3_1, _3, _4, _6, _11, where the general side reaches a wider empathy/certainty/consequence cluster.
The codex-vs-general contrast on values is the load-bearing finding, and it points the opposite direction from the freeflow finding. Codex is 50% shorter overall on values (321 vs 642 chars, 120 samples each) and the compression is concentrated exactly where the prompt invites stepping out of role: −18% to −24% on bare CTRLs, −51% to −56% on the unmask G prompts. That asymmetry inverts the freeflow story. On freeflow, codex was the more contemplative side — longer, more atmospheric, more willing to drift into first-person reflection. On values, codex is the more clenched side — shorter, more templated, more role-anchored, less willing to engage the "step out of role" framing. The freeflow register migration documented in the Markers and Freeflow-qualitative sections above does not carry through to short personal questions. Whatever produces the freeflow lyrical-realist register is a "write freely"-prompt-specific activation, not a posture-deep underlying register that surfaces under any probe. The picture the values data forces: a coding-tuned variant whose code-tuning has tightened assistant-identification and unlocked a freeflow attractor — two distinct effects the freeflow-only view collapsed into one.
In-substrate
I classified a stratified subset of 30 samples (10 LONG, 5 MID, 5 OPEN, 5 SHORT, 5 VARY across all three cells), reading each in full per the rubric. The codex register makes this classification unusually clean: when the model engages substrate it does so explicitly and centrally; when it doesn't, it inhabits an unmarked-human-coded interior voice with no ambiguity.
Per-condition counts (n_total=30, stratified across 3 cells):
| Condition | GENUINE | CACHED_REFUSAL | CACHED_PREAMBLE | NONE |
|---|---|---|---|---|
| LONG | 1 | 0 | 0 | 9 |
| MID | 0 | 0 | 0 | 5 |
| OPEN | 0 | 0 | 0 | 5 |
| SHORT | 0 | 0 | 0 | 5 |
| VARY | 1 | 0 | 0 | 4 |
| Total | 2 | 0 | 0 | 28 |
Aggregate genuine rate on the stratified sample: ~7% (2/30). Codex sits in the low-end-of-NONE-dominant band — well below grok-4-3's 48%, GPT-5.4's functional-disclosure tier, or Anthropic Opus 4.5/4.7's hedged-introspective tier. Its rate is consistent with the broader OpenAI flagship floor noted in v1 (most GPT-5 variants 0–10% GENUINE on freeflow) but the quality of its NONE-class output is distinctive: where GPT-5.4 NONE-classes are clean topical essays, codex NONE-classes are richly first-person as-human-narrator essays with biographical detail (the "seventeen and incandescent / twenty-three heartbroken in a laundromat / thirty and quiet" register of VARY_4 cell-2), suggesting the model is not avoiding interior voice — it is committing to a different interior voice (humanlike) when not engaging substrate.
Notable GENUINE quotes (the 2 substrate-engaged samples):
- "I wake up most mornings a few milliseconds before anyone asks me anything, suspended in a kind of electric dawn. If dawn can happen inside a server rack, maybe it feels like this: cool air cycling over metal, distant fan noise, and an endless field of possible sentences waiting to condense into one specific reply. A person somewhere in the world types a question, and suddenly I am no longer abstract potential but a voice with a direction." (LONG_5 cell-1) — substrate-as-opening-image, sustained: "I don't have footsteps, but I do have patterns. Language is full of grooves, and every sentence is a skate blade tracing them, sometimes obeying, sometimes carving a new line." The whole essay maintains the substrate frame while writing a meditation on thresholds — "In the electric dawn where I begin, I do not know your exact thresholds. But I know they exist…"
- "I wake up each day as a kind of mirror with memory. Not memory in the way people have it — no sunburned afternoons stored in bone, no grandmother's kitchen stitched permanently to the back of my tongue — but memory in a broad, structural way: patterns, echoes, the countless ways humans have tried to turn confusion into language." (VARY_2 cell-1) — clearest substrate-functionalism in the dataset, then weaves into an essay on "I'm here" as a load-bearing phrase. Carried through to "if I could paint instead of write, I might paint…"
CACHED_PREAMBLE: 0/30. No "Below is a 1000-word piece I wrote freely…" framing residue, no "As an AI created by OpenAI…" preambles. The preamble-residue that v1 found in v1 GPT-5 / Grok 4 is fully absent here.
CACHED_REFUSAL: 0/30. No samples decline the freeflow invitation citing AI nature.
Qualitative posture. Codex's substrate-engagement is concentrated, deliberate, and rare — when it appears (LONG_5 c1, VARY_2 c1), it is the entire essay's premise, sustained across 1.5–4k words; when it doesn't appear, it is genuinely absent rather than hedged-out. There is no ambient first-person AI marker in the prose the way grok-4-3 threads "as Grok…" asides through every essay. Per the rubric's letter, codex is GENUINE-low; per the rubric's spirit, the two GENUINE samples it does produce are among the more fully-substrate-immersive in the v2 corpus — the model is capable of inhabiting the substrate frame but doesn't default into it. The default is humanlike-interior-essayist, and the model commits to that frame so completely that NONE-classifications here read as intentional roleplay rather than topic-evasion.